This PDF file contains the front matter associated with SPIE Proceedings Volume 10576, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and Conference Committee listing.

Using diffusion tensor imaging (DTI), we developed and validated an automated classification procedure for Alzheimer’s disease (AD); specifically, DTI-derived fractional anisotropy (FA) and trace images from 22 AD subjects and 15 healthy control (HC) subjects were used. A total of four types of region of interest (ROI)-based features were tested, including the probability distribution distances of FA and trace images, within each of 162 whole-brain segmented ROIs, under both discrete and continuous intensity distribution modeling. The continuous modeling was conducted through a mixture of Gaussians, the parameters of which were estimated using maximum likelihood estimation via the expectation-maximization algorithm. We used principal component analysis (PCA) to reduce the dimension of the feature space and then linear discriminant analysis and support vector machine (SVM) for automated classification. According to our 10-times 10-fold cross-validation experiments, using the combination of PCA and linear SVM, the continuous distance of the trace image yielded the best classification performance with the accuracy being 87.84%±3.43% and the area under the receiver operating characteristic curve being 0.9121±0.0176, indicating its great potential as an effective AD biomarker.

Traditionally, brain function analysis is based on the magnitude data of complex-valued spatiotemporal (4D) functional magnetic resonance imaging (fMRI). Since an MRI signal is formed from the underlying brain tissue magnetic property through a cascade of transformations (such as dipole magnetization), the fMRI data (either magnitude or phase) do not directly capture the original magnetic source. In principle, upon solving an inverse fMRI problem, we can reconstruct the magnetic source (specifically magnetic susceptibility, denoted by χ and analyze brain function in the reconstructed χ dataspace at a stage closer to the origin of brain function neurophysiology. Our recent research has shown that the magnetic χ source can be reconstructed from the fMRI phase through a computational inverse MRI solution (CIMRI). Together with the fMRI output data, we can compare three aspects of the data, the magnitude, the phase, and the susceptibility, each of which provides a different perspective. Given a 4D dataset, we analyze the data via independent component analysis (ICA), applicable to both single-subject and multi-subject data. In this study, we addressed the following points: 1) brain function ICA decomposition of magnitude (mICA), phase (pICA), and susceptibility (χICA); 2) comparison of brain function network connectivity matrices (FC) for each of these, namely {mFC, pFC, and χFC} matrices; and 3) applications to a task fMRI experiment (fingertapping, 20 subjects). In theory, we show that the fMRI phase is approximately linearly related to the reconstructed χ source data (different by a spatial dipole convolution), while fMRI magnitude has a nonlinear relationship. Therefore, we conclude that pFC is more similar to χFC than mFC. Through experimental data analyses, we have verified this conclusion.

A key feature of magnetic resonance (MR) imaging is its ability to manipulate how the intrinsic tissue parameters of the anatomy ultimately contribute to the contrast properties of the final, acquired image. This flexibility, however, can lead to substantial challenges for segmentation algorithms, particularly supervised methods. These methods require atlases or training data, which are composed of MR image and labeled image pairs. In most cases, the training data are obtained with a fixed acquisition protocol, leading to suboptimal performance when an input data set that requires segmentation has differing contrast properties. This drawback is increasingly significant with the recent movement towards multi-center research studies involving multiple scanners and acquisition protocols. In this work, we propose a new framework for supervised segmentation approaches that is robust to contrast differences between the training MR image and the input image. Our approach uses a generative simulation model within the segmentation process to compensate for the contrast differences. We allow the contrast of the MR image in the training data to vary by simulating a new contrast from the corresponding label image. The model parameters are optimized by a cost function measuring the consistency between the input MR image and its simulation based on a current estimate of the segmentation labels. We provide a proof of concept of this approach by combining a supervised classifier with a simple simulation model, and apply the resulting algorithm to synthetic images and actual MR images.

Background: Clustering thalamic nuclei is important for both research and clinical purposes. For example, ventral intermediate nuclei in thalami serve as targets in both deep brain stimulation neurosurgery and radiosurgery for treating patients suffering from movement disorders (e.g., Parkinson's disease and essential tremor). Diffusion magnetic resonance imaging (dMRI) is able to reflect tissue microstructure in the central nervous system via fitting different models, such as, the diffusion tensor (DT), constrained spherical deconvolution (CSD), neurite orientation dispersion and density imaging (NODDI), diffusion kurtosis imaging (DKI) and the spherical mean technique (SMT). Purpose: To test which of the above-mentioned dMRI models is better for thalamic parcellation, we proposed a framework of k-means clustering, implemented it on each model, and evaluated the agreement with histology. Method: An ex vivo monkey brain was scanned in a 9.4T MRI scanner at 0.3mm resolution with b values of 3000, 6000, 9000 and 12000 s/mm². K-means clustering on each thalamus was implemented using maps of dMRI models fitted to the same data. Meanwhile, histological nuclei were identified by AChE and Nissl stains of the same brain. Overall agreement rate and agreement rate for each nucleus were calculated between clustering and histology. Sixteen thalamic nuclei on each hemisphere were included. Results: Clustering with the DKI model has slightly higher overall agreement rate but clustering with other dMRI models result in higher agreement rate in some nuclei. Conclusion: dMRl models should be carefully selected to better parcellate the thalamus, depending on the specific purpose of the parcellation.

The lack of standardizable objective diagnostic measurement techniques is a major hurdle in the assessment and treatment of pediatric patients with thoracic insufficiency syndrome (TIS). The aim of this paper is to explore quantitative dynamic MRI (QdMRI) volumetric parameters derived from thoracic dMRI in pediatric patients with TIS and the relationships between dMRI parameters and clinical measurements. 25 TIS patients treated with vertical expandable prosthetic titanium rib (VEPTR) surgery are included in this retrospective study. Left and right lungs at endinspiration and end-expiration are segmented from constructed 4D dMRI images. Lung volumes and excursion (or tidal) volumes of the left/right chest wall and hemi-diaphragms are computed. Commonly used clinical parameters include thoracic and lumbar Cobb angles and respiratory measurements from pulmonary function testing (PFT). 200 3D lungs in total (left & right, pre-operative & post-operative, end-inspiration & end-expiration) are segmented for analysis. Our analysis indicates that change of resting breathing rate (RR) following surgery is negatively correlated with that of QdMRI parameters. Chest wall tidal volumes and hemi-diaphragm tidal volumes increase significantly following surgery. Clinical parameter RR reduced after surgical treatment with P values around 0.06 but no significant differences were found on other clinical parameters. The significant increase in post-operative tidal volumes suggests a treatment-related improvement in lung capacity. The reduction of RR following surgery shows that breathing function is improved. The QdMRI parameters may offer an objective marker set for studying TIS, which is currently lacking.

Purpose: Time-of-flight (TOF) been successfully implemented in whole body PET, significantly improving clinical performance. However, TOF has not been a priority in development of dedicated brain PET systems due the relatively small size of the human head, where coincidence timing resolution (CTR) below 200 ps is necessary to arrive at substantial performance improvements. The Brain PET (BET) consortium is developing a PET detector block with ultrafast CTR, high sensitivity and high spatial resolution (X, Y, depth of interaction, DOI) that provides a pathway to significantly improved brain PET. Methods: We have implemented analytical and Monte Carlo models of scintillation photons transport in scintillator segments with the trans-axial cross-section equal or smaller than 3x3 mm² . Results: The signal amplitude and timing of W mm x W mm x L mm scintillators (1 mm<W<3 mm, 5 mm <L< 30 mm) are strongly influenced by sidewall surface polish and external reflector. Highly polished surfaces provide nearly perfect total internal reflection (TIR), enabling the ultrafast timing performance to be relatively independent of scintillator crosssection. The signal amplitude in such a configuration does not depend on DOI. However, the differential signal from top and bottom SiPM in the dual-ended readout can be used to determine DOI. Using TIR alone, the average of the photon detection times at the top and bottom SiPMs provides a good estimation of the gamma ray absorption time. Averaging ~10 photons starting from 3^rd photon produces the shortest CTR for SPTR=50 ps. Conclusions: We established that the advanced silicon photomultiplier designs with high single photon detection efficiency (QE=60%) and high single photon timing resolution (SPTR =50 ps) are critical for achieving ultrafast TOF-PET performance with CTR ~50 ps and ~4 mm DOI resolution.

Techniques for iterative reconstruction of magnetic resonance diffusion images from motion corrected multi-planar acquisitions are beginning to allow the use of more complex diffusion models and tractography techniques to study early brain connectivity. Many techniques have been developed for adult and neonatal brain tractography from diffusion images. However, fundamental differences in the underlying tissues, signal levels and relative spatial resolution that are available in fetal studies mean these techniques may need to be significantly adapted to deal with the different challenges. Here we evaluate and compare the use of diffusion tensor and spherical harmonic models in extracting known fetal white matter connective anatomy from multi-planar, motion corrected, variable data density studies of normally developing human fetal brains. Visual evaluation of known tracts indicates that, although there are significant differences in the diffusion properties of fetal brain tracts and also image signal strength in fetal brain studies, when compared to adult brain imaging and tractography, high order models such as spherical harmonic techniques still offer advantages in appropriately delineating known anatomy from in utero data.

The study of functional connectivity of the human brain has provided valuable insights into its organization principles. Studies have revealed consistent and reproducible patterns of activity across individuals which are referred to as resting-state networks. Although these have been studied extensively, the direction of information flow between these regions is less understood. We aimed to study this by analyzing resting state scans from 20 subjects (11 male and 9 female, all healthy) and capturing the functional interdependence of 32 regions of interest spanning the different resting state networks using a Mutual Connectivity Analysis (MCA) framework with non-linear time series modeling based on Generalized Radial Basis function (GRBF) neural networks. The resulting networks are then analyzed to explore patterns of directed connectivity across the subjects. Using the general linear model, we observe that the nodes of the salience network particularly shows patterns of directed influence within as well as outside the network (p<0.05, FDR corrected). Additionally, the anterior cingulate cortex exhibits a strong outgoing influence on various regions of the brain. Such directional influences of the RSNs have not been reported previously. These results suggest that our framework can effectively capture patterns of distributed and directed connectivity occurring in the brain network and can therefore serve to enhance our understanding of its organizational principles.

The advent of advanced multivariate time-series analysis methods to capture directional information flow in the brain, such as large-scale Granger causality (lsGC), has already yielded insights into healthy and diseased brain states and holds promise for continued discoveries in neuroscience. Here, a set of functional brain networks was generated by applying lsGC to a resting-state functional MRI dataset comprised of 20 healthy individuals in order to characterize network properties across a range of spatial scales and network densities. Network vertex definition was performed using brain parcellation templates at three spatial resolutions (90, 231, and 438 regions of interest) and eight network densities. The small-worldness parameter σ and degree-distribution were computed for each network. Over all thresholds and spatial resolutions considered, the networks were determined to have small-world properties, as indicated by mean σ » 1. Moreover, all networks were determined to have an exponentially truncated degree distribution, indicating a network structure highly resilient to both hub and random vertex attack. In addition, k-nearest neighbors thresholding and Louvain community detection was performed on the mean network from all subjects to extract functional modules. Eight functional modules were discovered, three of which corresponded to known resting-state networks, including the default mode, visual, and sensorimotor networks. The results of these analyses show that lsGC has the ability to capture important aspects of brain connectivity, including small-worldness, resilience to attack, and functional modularity.

Large-scale Granger causality (lsGC) analysis quantifies multivariate voxel-resolution connectivity in resting-state functional MRI (fMRI) unlike commonly used multivariate approaches that estimate connectivity at a coarse resolution. We investigate the effect noise and repetition time (TR) of fMRI signals have on the ability of lsGC to capture true connectivity and compare with traditionally used multivariate Granger causality analysis (mvGC). To this end, we use realistic fMRI simulations, generated with varying TR and noise levels, for fifty-node simulations. LsGC produces directed connectivity graphs, represented as connectivity matrices which we compare with the known ground truth of the simulations with the Area Under the receiver operating characteristic Curve (AUC) as a measure of agreement. The best AUC with lsGC was 0.957 while the least was 0.835 at TR = 3 s. Our results show that lsGC performs much better than mvGC approaches for both noise levels and different TR. An interesting finding with lsGC was that at higher sampling rate, corresponding to TR < 2 s increase in noise did not significantly reduce performance. However, as with increasing TR beyond 2 s, the effects of noise in the system is no longer negligible. Our results indicate that if the TR is sufficiently small, the performance of lsGC is not hindered greatly by noise levels. However, at higher TR, the deterioration of performance due to high TR is compounded by higher noise levels, indicating that improvements in TR may be more beneficial in extracting accurate lsGC connections.

The cerebellum plays an important role in both motor control and cognitive functions. Several methods to automatically segment different regions of the cerebellum have been recently proposed. Usually, the performance of the segmentation algorithms is evaluated by comparing with expert delineations. However, this is a laboratory approach and is not applicable in real scenarios where expert delineations are not available. In this paper, we propose a method that can automatically detect cerebellar lobule segmentation outliers. Instead of only evaluating the final segmentation result, the intermediate output of each segmentation step is evaluated and considered using a Hidden Markov Model (HMM) to produce a global segmentation assessment. For each intermediate step, a state-of-the-art image classification model Bag-of-Words" (BoW) is applied to quantize features of segmentation results, which then serves as observations of the trained HMM. Experiments show that the proposed method achieves both a high accuracy on predicting Dice of upcoming segmentation steps, and a high sensitivity to outlier detection.

We present new methods for segmentation of hippocampal dendritic spines in 3D confocal light microscopy and computation of 3-D morphological attributes characterizing dendritic spine plasticity. The methods are applied on 3D confocal light microscopy images of dendritic spines from dissociated hippocampal cultures. The segmentation method is based on the principle of multi-scale opening which uses a set of user-specified seeds and iteratively segments structures at different scales starting at a large scale and progressing toward finer scales. The accuracy of the segmentation method is evaluated by comparing its results with the gold standard obtained by manual labelling. The reproducibility of the overall method involving segmentation as well as computation of structural measures is assessed by comparing the values of structural measures derived from segmentation results generated using seeds from three mutually-blinded users. Finally, the performance of the overall method is examined in terms of its ability to characterize spine morphological changes after chemically induced long-term potentiation.

Normal pressure hydrocephalus (NPH) is a brain disorder caused by disruption of the flow of cerebrospinal fluid (CSF). The dementia-like symptoms of NPH are often mistakenly attributed to Alzheimer’s disease. However, if correctly diagnosed, NPH patients can potentially be treated and their symptoms reversed through surgery. Observing the dilated ventricles through magnetic resonance imaging (MRI) is one element in diagnosing NPH. Diagnostic accuracy therefore benefits from accurate, automatic parcellation of the ventricular system into its sub-compartments. We present an improvement to a whole brain segmentation approach designed for subjects with enlarged and deformed ventricles. Our method incorporates an adaptive ventricle atlas from an NPH-atlas-based segmentation as a prior and uses a more robust relaxation scheme for the multi-atlas label fusion approach that accurately labels the four sub-compartments of the ventricular system. We validated our method on NPH patients, demonstrating improvement over state-of-the-art segmentation techniques.

Emerging evidence suggests that the lymphatics play an important role in the pathogenesis and progression of peripheral arterial and venous diseases. In two pilot studies, we sought to evaluate lymphatic function of patients with (i) late stage chronic venous disease (CVD) with active venous leg ulcers (VLUs) and (ii) early CVD and mild to moderate peripheral arterial disease (PAD) using near‐infrared fluorescence lymphatic imaging (NIRFLI). After informed consent and under an FDA-approved IND for the off‐label administration of indocyanine green (ICG), each subjects received intradermal injections of ICG on the feet and legs. Imaging of lymphatic anatomy and pumping activity was performed by illuminating the legs with near-infrared light and collecting the resultant fluorescent light emanating from the ICG‐laden lymph using a custom imaging system. Data collection occurred in an outpatient setting between the dates of 2009 until present, with the study of early CVD and PAD subjects underway. We observed abnormal lymphatic anatomy and reduced lymphatic pumping in all subjects enrolled in these two pilot studies. Observed abnormal lymphatic anatomy, as compared to previously imaged healthy subjects, included dermal lymphatic backflow as well as segmented, dilated and/or tortuous lymphatic vessels. Reduced lymphatic pumping was also observed in all subjects, and lymphatic reflux was noted in those subjects with an arterial component to their disease. While these studies continue, evidence is mounting that lymphatic dysfunction is associated with the etiology of peripheral arterial and venous diseases. Supported in parts by the National Institutes of Health R21 HL132598‐01 and Tactile Medical.

We present a method to estimate myocardial blood flow confidence intervals (MBF-CI) for model-based analysis of CT myocardial perfusion imaging (CT-MPI). We have determined that good fits, as assessed with visual evaluation, root-mean-square error, and Akaike information criterion (AIC), can lead to very poor MBF estimates with >50% error. We develop the use of confidence intervals to help confirm that good models are leading to good MBF estimates. We assess MBF precision for multiple analysis models from the literature, including adiabatic approximation of tissue homogeneity (AATH), plasma tissue uptake (PTU), and a newly proposed robust physiologic model (RPM). For evaluation, we use a physiologic simulator, digital CT-MPI phantom, and in vivo CT-MPI data from a porcine model of coronary stenosis. MBF-CI was calculated using empirical likelihood to determine the range of MBF values that fall within the 95% joint parameter confidence region. On simulated data, although AIC was smallest (preferred) for AATH and greatest for RPM, standard deviation of MBF measurements was between 7-41 times greater for AATH than RPM, indicating RPM significantly improved MBF precision. MBF-CI appropriately selected RPM for best MBF precision. For the SNR=20 example condition, standard deviations were 1.7, 28.4, and 34.7mL/min/100g; MBF-CIs were 26, 375, and 435mL/min/100g; and AICs were 299.7, 253.4, and 245.3 for RPM, PTU, and AATH, respectively. Overall, best MBF precision was ranked RPM>PTU>AATH. These findings suggest that models with fewer free parameters, such as RPM, yield precise MBF measurements and that MBF-CI can select for models with good MBF measurement precision.

The overall geometry and different biomechanical parameters of an abdominal aortic aneurysm (AAA), contribute to its severity and risk of rupture, therefore they could be used to track its progression. Previous and ongoing research efforts have resorted to using uniform material properties to model the behavior of AAA. However, it has been recently illustrated that different regions of the AAA wall exhibit different behavior due to the effect of the biological activities in the metalloproteinase matrix that makes up the wall at the aneurysm site. In this work, we introduce a non-invasive patientspecific regional material property model to help us better understand and investigate the AAA wall stress distribution, peak wall stress (PWS) severity, and potential rupture risk. Our results indicate that the PWS and the overall wall stress distribution predicted using the proposed regional material property model, are higher than those predicted using the traditional homogeneous, hyper-elastic model (p <1.43E-07). Our results also show that to investigate AAA, the overall geometry, presence of intra-luminal thrombus (ILT), and loading condition in a patient specific manner may be critical for capturing the biomechanical complexity of AAAs.

Purpose: 3D printed patient specific vascular models provide the ability to perform precise and repeatable benchtop experiments with simulated physiological blood flow conditions. This approach can be applied to CT-derived patient geometries to determine coronary flow related parameters such as Fractional Flow Reserve (FFR). To demonstrate the utility of this approach we compared bench-top results with non-invasive CT-derived FFR software based on a computational fluid dynamics algorithm and catheter based FFR measurements.

Materials and Methods: Twelve patients for whom catheter angiography was clinically indicated signed written informed consent to CT Angiography (CTA) before their standard care that included coronary angiography (ICA) and conventional FFR (Angio-FFR). The research CTA was used first to determine CT-derived FFR (Vital Images) and second to generate patient specific 3D printed models of the aortic root and three main coronary arteries that were connected to a programmable pulsatile pump. Benchtop FFR was derived from pressures measured proximal and distal to coronary stenosis using pressure transducers.

Results: All 12 patients completed the clinical study without any complication, and the three FFR techniques (Angio-FFR, CT-FFR, and Benchtop FFR) are reported for one or two main coronary arteries. The Pearson correlation among Benchtop FFR/ Angio-FFR, CT-FFR/ Benchtop FFR, and CT-FFR/ Angio-FFR are 0.871, 0.877, and 0.927 respectively.

Conclusions: 3D printed patient specific cardiovascular models successfully simulated hyperemic blood flow conditions, matching invasive Angio-FFR measurements. This benchtop flow system could be used to validate CTderived FFR diagnostic software, alleviating both cost and risk during invasive procedures.

Myocardial scar geometry may influence the sensitivity of predicting risk for ventricular tachycardia (VT) using computational models of the heart. This study aims to compare the differences in reconstructed geometry of scar generated using two-dimensional (2D) versus three-dimensional (3D) late gadolinium-enhanced magnetic resonance (LGE-MR) images. We used a retrospectively-acquired dataset of 17 patients with myocardial scar who underwent both 2D and 3D LGE-MR imaging. We segmented the scar manually in both 2D and 3D LGE-MRI using a multi-planar image processing software. We then reconstructed the 2D scar segmentation boundaries from 2D LGE-MRI to 3D surfaces using a LogOdds-based interpolation method. Finally, we assessed the 3D models of scar in both 3D and 2D-reconstructed techniques using several shape and volume metrics such as, fractal dimensions, number of connected components, mean scar volume, and normalized scar volume. The higher fractal dimension resulted for 3D may indicate that the 3D LGE-MRI produces a more complex surface geometry by better capturing the intact geometry of the scar. The 2D LGE-MRI produced a larger normalized scar volume (19.48±10 cm³) than the 3D LGE-MRI (10.92±7.12 cm³). We also provided a statistical analysis on the scar volume differences acquired from 2D and 3D LGE-MRI.

The imaging of endovascular devices during neurovascular procedures such as the coiling of aneurysms guided with CBCT imaging may be challenging due to the presence of highly attenuating materials such as platinum in the coil and stent marker, nickel-titanium in the stent, iodine in the contrast agent, and tantalum in the embolization agent. The use of dual-energy imaging followed by a basis material decomposition image processing-scheme may improve the feature separation and recognition. Two sets of testing were performed to validate this concept. The first trial was the acquisition of dual-energy micro-CBCT data of a 3D-printed simple aneurysm model using a 49.5 μm pixel size CMOS detector (Teledyne DALSA, Waterloo, ON.). Two sets of projection data were acquired using beam energies of 35 kVp and 70 kVp. Axial slices were reconstructed and used to carry out the material decomposition processing. The second trial was the acquisition of dual-energy CBCT images of a RS-240T angiographic head phantom (Radiology Support Devices Inc., CA.) with an iodine vascular insert using a Toshiba Infinix BiPlane C-arm system coupled to a flat panel detector. Two sets of image data were acquired using beam energies of 80 kVp and 120 kVp. Following image reconstruction, slices of the phantom were decomposed using the same processing as previously. The resulting image data over both trials indicate that the decomposition process was successful in separating the kinds of materials commonly used during a neurovascular intervention, such as platinum, cobalt-chromium, and iodine. The normalized root mean square error metric was used to quantitatively assess this. This indicates a basis for future more clinically relevant testing of our methods.

Super-resolution ultrasound (SR-US) imaging can achieve a ten-fold resolution improvement compared with the traditional ultrasound technique, which is important for the medical diagnosis and treatment. However, challenges remain in SR-US imaging. In this paper, on one hand, a Gaussian fitting method, derived from optical localization microscopy, is used to improve the imaging spatial resolution of the SR-US. On the other hand, a plane wave technique is also used in US imaging for improving the imaging speed of the SR-US. To evaluate the performance of the proposed method, the numerical simulation was performed based on a phantom model. The experimental results indicate that by the use of a Gaussian fitting location method, combined with a plane wave transmission technique, we can accurately image the movement of microbubble in the phantom at a high frame rate, compared to the conventional B-model imaging. Hence, the technique makes it possible to achieve fast SR-US imaging.

Modern 3D printing technology allows rapid prototyping of vascular phantoms based on an actual human patient with a high degree of precision. Using this technology, we present a platform to accurately simulate clinical views of neuro-endovascular interventions and devices. The neuro-endovascular interventional phantom has a 3D printed cerebrovasculature model derived from a patient CT angiogram and embedded inside a human skull providing bone attenuation. Acrylic layers were placed underneath and on top of the skull, simulating entrance and exit tissue attenuation and also simulating forward scatter.

The 3D model was connected to a pulsatile flow loop for simulating interventions using clinical devices such as catheters and stents. To validate the x-ray attenuation and establish clinical accuracy, the automatic exposure selection by a clinical c-arm system for the phantom was compared with that for a commercial anthropomorphic head phantom (SK-150, Phantom Labs). The percentage difference between automatic exposure selection for the neuro-intervention phantom and the SK-150 phantom was under 10%.

By changing 3D printed models, various patient diseased anatomies can be simulated accurately with the necessary x-ray attenuation. Using this platform various interventional procedures were performed using new imaging technologies such as a high-resolution x-ray fluoroscope and a dose-reduced region-of-interest attenuator and differential temporally filtered display for enhanced interventional imaging. Simulated clinical views from such phantom-based procedures were used to evaluate the potential clinical performance of such new technologies.

Ultrasound (US) imaging technique is currently one of the most common imaging techniques in clinical application, but the spatial resolution is low. Recently, with the aid of contrast agents, super-resolution ultrasound imaging technique has been proposed, which can overcome the diffraction limit in US by using the super-localization method, similar to superresolution optical microscopy. But, there is still a trade-off between spatial and temporal resolution in super-resolution US imaging. To address the problem, inspired by super-resolution optical fluctuation imaging (SOFI), in this paper, we apply SOFI to US imaging to achieve a good imaging performance. Further, to cancel the nonlinear response to brightness in SOFI, the balanced SOFI (bSOFI) is also used in this paper, which allows to achieve the higher spatial resolution. To evaluate the feasibility of the proposed method, the numerical simulation was performed based on a dynamic phantom model, which was scanned by synthetic transmit aperture (STA) technique. The result indicates that by using the proposed method (SOFI or bSOFI), the imaging performance of US can be improved compared to STA. In addition, when using bSOFI method, the imaging performance of super-resolution US can be further improved, compared with SOFI method.

To reduce the artifacts and improve the image quality in sparse-view CT reconstruction. A novel improved GoogLeNet is proposed to reduce artifacts of the sparse-view CT reconstruction. This paper uses the residual learning for GoogLeNet to study the artifacts of sparse-view CT reconstruction, and then subtracts the artifacts obtained by learning from the sparse reconstructed images, finally recovers a clear correction image. The intensity of reconstruction using the proposed method is very close to the full-view projective image. The results indicate that the proposed method is practical and effective for reducing the artifacts and preserving the quality of the reconstructed image.

To date, there is no single sign, symptom, or test that can clearly diagnose early stages of Temporomandibular Joint Osteoarthritis (TMJ OA). However, it has been observed that changes in the bone occur in early stages of this disease, involving structural changes both in the texture and morphometry of the bone marrow and the subchondral cortical plate. In this paper we present a tool to detect and highlight subtle variations in subchondral bone structure obtained from high resolution Cone Beam Computed Tomography (hr-CBCT) in order to help with detecting early TMJ OA. The proposed tool was developed in ITK and 3DSlicer and it has been disseminated as open-source software tools. We have validated both our texture analysis and morphometry analysis biomarkers for detection of TMJ OA comparing hr-CBCT to μCT. Our initial statistical results using the multidimensional features computed with our tool indicate that it is possible to classify areas of demonstrated loss of trabecular bone in both μCT and hr-CBCT. This paper describes the first steps to alleviate the current inability of radiological changes to diagnose TMJ OA before morphological changes are too advanced by quantifying subchondral bone biomarkers. This paper indicates that texture based and morphometry based biomarkers have the potential to identify OA patients at risk for further bone destruction.

Cancer development in oral epithelial tissue induces subtle changes in tissue autofluorescence that are associated with increased metabolic activity in malignant oral epithelial cells. These autofluorescence biomarkers of oral cancer progression include a decrease in the optical “redox ratio”, defined as the autofluorescence intensity of NADH divided by that of FAD, and specific changes in the fluorescence lifetime of both NADH and FAD. We therefore hypothesized that more specific biomarkers of oral cancer and dysplasia can more accurately be quantified by endogenous fluorescence lifetime imaging (FLIM). In this work, FLIM images of benign, dysplastic and early stage cancerous oral lesions from 52 patients were acquired at three emission channels (390±20nm, 452±22.5nm and >500nm) using a handheld multispectral FLIM endoscope. For each pixel, the fluorescence decays collected at the three emission bands were analyzed using a biexponential decay model, resulting on 16 FLIM-derived parameters per pixel, which generated multiparametric FLIM images of each oral lesion. Statistical analysis was performed on each of the computed FLIM parameters (Wilcoxon test: Normal vs. Benign, Normal vs. Dysplasia/Cancer; Mann-Whitney test: Benign vs. Dysplasia/Cancer). Results from this analysis revealed that FLIM-derived parameters associated with collagen lifetime, NADH lifetime, FAD autofluorescence, and the optical redox ratio were statistically different between dysplastic/cancerous vs. benign oral lesions. This study provides the first demonstration for the clinical imaging of autofluorescence biochemical and metabolic biomarkers of oral epithelial cancer and dysplasia, which could potentially enable early detection of oral cancer.

The pulvinar of the thalamus is a higher-order thalamic nucleus that is responsible for gating information flow to the cortical regions of the brain. It is involved in several cortico-thalamocortical relay circuits and is known to be affected in a number of neurological disorders. Segmenting the pulvinar in clinically acquired images is important to support studies exploring its role in brain function. In recent years, we have proposed an active shape model method to segment multiple thalamic nuclei, including the pulvinar. The model was created by manual delineation of high resolution 7T images and the process was guided by the Morel stereotactic atlas. However, this model is based on a small library of healthy subjects, and it is important to validate the reliability of the segmentation method on a larger population of clinically acquired images. The pulvinar is known to have particularly strong white matter connections to the hippocampus, which allows us to identify the pulvinar from thalamic regions of high hippocampal structural connectivity. In this study, we obtained T1-weighted and diffusion MR data from 43 healthy volunteers using a clinical 3T MRI scanner. We applied the segmentation method to the T1-weighted images to obtain the intrathalamic nuclei, and we calculated the connectivity maps between the hippocampus and thalamus using the diffusion images. Our results show that the shape model segmentation consistently localizes the pulvinar in the region with the highest hippocampal connectivity. The proposed method can be extended to other nuclei to further validate our segmentation method.

Myocardial scar, a non-viable tissue which forms in the myocardium due to insufficient blood supply to the heart muscle, is one of the leading causes of life-threatening heart disorders, including arrhythmias. Accurate reconstruction of myocardial scar geometry is important for diagnosis and clinicial prognosis of the patients with ischemic cardiomyopathy. The 3D late gadolinium-enhanced magnetic resonance imaging (LGE-MRI) is increasingly being investigated for assessing myocardial tissue viability. For applications, such as computational modeling of cardiac electrophysiology aimed at stratifying patient risk for post-infarction arrhythmias, segmentation and reconstruction of the intact geometry of scar is required. However, manual analysis and segmentation of myocardial scar from 3D LGE-MRI is a tedious task. Therefore, semi-automated and fully-automated segmentation algorithms are highly desirable in a clinical setting. In this study, we developed an approach to segment the myocardial scar from 3D LGE-MR images using a continuous max-flow (CMF) method. The data term comprised of a distribution matching term for scar and normal myocardium and a boundary smoothness term for the scar boundaries. The region-of-interest for the scar segmentation is constrained, using manually segmented myocardium. We evaluated our CMF method for accuracy by comparing it to manual scar delineations using 3D LGE-MR images of 34 patients. We compare the results of the CMF technique to ones by conventional full-width-at-half-maximum (FWHM) and signal-threshold-to-reference-mean (STRM) methods. The CMF method yielded a Dice similarity coefficient (DSC) of 72±18% and an absolute volume error (|V E|) of 15.42±14.1 cm³. Overall, the CMF method outperformed the state-of-the-art methods for all reported metrics in 3D scar segmentation except for the recall value which STRM 2-SD performed better than CMF on average.

A surgical simulator with elaborate artificial eyeball models has been developed for ophthalmic surgeries, in which sophisticated skills are required. To create the elaborate eyeball models with microstructures included in an eyeball, a database of eyeball models should be compiled by segmenting eye structures based on high-resolution medical images. Therefore, this paper presents an automated segmentation of eye structures from micro-CT images by using Fully Convolutional Networks (FCNs). In particular, we aim to construct a method for accurately segmenting eye structures from sparse annotation data. This method performs end-to-end segmentation of eye structures, including a workflow from training the FCN based on sparse annotation to obtaining the segmentation of the entire eyeball. We use the FCN trained on the slices sparsely annotated in a micro-CT volume to segment the remaining slices in the same volume. To achieve accurate segmentation from less annotated images, the multi-class segmentation is performed by using the network trained on the preprocessed and augmented micro-CT images; in the preprocessing, we apply filters for removing ring artifacts and random noises to the images, while in the data augmentation process, rotation and elastic deformation operations are performed on the sparsely-annotated training data. From the results of experiments for evaluating segmentation performances based on sparse annotation, we found that the FCN trained with data augmentation could achieve high segmentation accuracy of more than 90% even from a sparse training subset of only 2.5% of all slices.

Head and neck (HN) cancer survivors are particularly susceptible to lymphedema with a reported incidence of 75%. Treatment of HN lymphedema (HNL) consists of manual lymphatic drainage and compression. In this FDA-approved study, we used near-infrared fluorescence lymphatic imaging (NIRFLI) to assess the response of HNL to a single session of sequential, pneumatic compression therapy (PCT) and after two weeks of daily PCT. Following informed consent, 5 HNL subjects were enrolled into this study which consisted of two imaging sessions where each subject received six intradermal injections of 25µg of ICG in 0.1ml saline. Immediately after injection, NIRFLI was performed by illuminating the HN with excitation light and collecting the resultant fluorescent signal using a custom imaging system. After initial imaging, each subject underwent a single PCT session after which NIRFLI was again performed. Subjects were provided a pneumatic compression device for daily use over two weeks, after which imaging was repeated. Abnormal dermal lymphatic backflow was observed in all subjects. Enhanced lymphatic uptake was observed in all subjects following a single PCT session. After two weeks of PCT, NIRFLI revealed the complete resolution of dermal backflow in one subject, the reduction of dermal backflow in two subjects, and no apparent lymphatic change in two subjects though both reported subjective improvements in swallowing or speaking. PCT therapy may be an effective method to treat HNL. Additional studies are needed to assess long term clinical and quality of life impact. This study was funded in part by Tactile Medical.

In this work, we compare standard wide-field fluorescence lifetime imaging microscopy (FLIM) and structured illumination FLIM (SI-FLIM) as methods for the early detection of oral squamous cell carcinoma (OSCC). Our technique, SI-FLIM, provides depth dependent fluorescence lifetime information of the oral epithelium, isolating the endogenous fluorophore of interest, NADH, from interfering fluorescence generated mainly by collagen in the lamina propria. Male golden Syrian hamsters (Cricetus auratus) were used as the animal model for OSCC. They were treated with a carcinogen, 7,12-Dimethylbenz[a]anthracene (DMBA), for a twelve-week period by applying the DMBA suspended in mineral oil to their cheek pouches 3 times per week. The progression of OSCC was monitored over a 12-week period with imaging beginning at the 6^th week. The cheek pouch with lesions was imaged in a 3x4 grid (twelve total images), with each section of the grid being correlated with histopathological analysis. The NADH fluorescence channel, as a diagnostic indicator, was compared for both SI-FLIM and widefield FLIM. ROC analysis, in the task of distinguishing between mild dysplasia and normal tissue, showed that SI-FLIM (AUC=0.83, se=0.07) may be a better indicator for early cases of mild dysplasia when compared to widefield FLIM (AUC=0.63, se=0.07) with statistical significance.

Glioma is one of the most important leading causes of cancer-related deaths worldwide. Temozolomide (TMZ) is a DNA methylating agent that presents promising antitumor activity against high grade glioma. However, there is no effective way to assess the therapeutic response to TMZ at early stage. In this study, we evaluated the efficacy of TMZ on brain tumor through bioluminescence tomography (BLT) based on multi-modality imaging system.

Initially, the human glioma cell line U87MG-fLuc cells were cultured, and the orthotopic mouse brain tumor model was established. 10 days after the tumor cell implantation, the mice were divided into two groups including the TMZ group and the control group. The mice in the TMZ group were treated with Temozolomide with dosage of 50 mg/kg/day intraperitoneally for continuous 6 days, and the mice in the control group were treated with sterile saline at equal volume. The bioluminescence imaging (BLI) was acquired every 5 days for monitoring the therapeutic responses. A randomly enhanced adaptive subspace pursuit (REASP) algorithm is presented for bioluminescence tomography reconstruction. Basically, numerical experiments were used to validate the efficiency of the proposed method, and then the mice’s CT and BLI data were acquired to reconstruct BLT using the REASP algorithm.

The results in this study showed that the growth of glioma can be monitored from very early stage, and the TMZ treatment efficacy can be reliably and objectively assessed using BLT method. Our data demonstrated TMZ can effectively inhibit the tumor growth.

Cone beam X-ray luminescence computed tomography (CB-XLCT) has recently been proposed as a new molecular imaging modality for various biomedical applications. It utilizes X-ray excitable nanophosphors to produce visible or near-infrared (NIR) luminescence and combines the high sensitivity of optical imaging with the high spatial resolution of X-ray imaging. With the development of the nanophosphors and reconstruction methods, dynamic XLCT imaging, which can reflect the dynamic course of absorption, distribution, and elimination of the nanophosphors in vivo, has demonstrated its initial prospect in biological and biochemical studies. However, challenges remain in resolving nanophosphors (drug) distributions inside the imaging object due to the high light scattering and complex dynamics of nanophosphor’s delivery. Considering that target with different functions may have different kinetic behaviors, in this paper we present a method to resolve targets with different kinetics by utilizing principal component analysis (PCA). The metabolic processes of nanophosphors (Y₂O₃:Eu³⁺) of two targets were simulated and imaged using a CB-XLCT system, with two targets located at different edge-to-edge distances of 0.12 cm. Simulation and experiment studies validate the performance of the proposed algorithm. The results suggest that two adjacent targets of different kinetic behaviors can be extracted and illustrated by the proposed method, at an edge-to-edge distance of 0.12 cm.

Fluorescence Molecular Tomography (FMT) is one of the most important preclinical research techniques, which can obtain three-dimensional reconstruction of tumors in mouse in vivo. However, the ill-posedness of FMT makes its reconstruction a challenging problem. Therefore, more effective, robust, and accurate reconstruction methods are needed to be developed to solve the FMT reconstruction problem.

In this paper, a reconstruction method named multipath subspace pursuit (MSP) is applied to solve the FMT problem. At the end of an iteration, the MSP method creates several candidate support set. Through evaluating the normal of final residual vector, the best candidate can be selected as the final support set. Then the support set is used for reconstructing sense matrix to achieve the goal of FMT reconstruction.

In order to verity the reconstruction result of the proposed MSP method, the simulated experiment of triple fluorescent sources and quantitative analyses of position error and relative intensity error for the experiment have been conducted. The MSP method obtains satisfactory results, and the source position error is below 1 mm. Moreover, the computation time of the MSP method is about one order of magnitude less than iterated shrinkage with the L1-norm (IS_L1) method. The MSP method not only can obtain the result of robustness but also can reduce the artifacts in the background. The above results revealed the MSP method for the potential FMT application.

Accurate and automated brain tumor segmentation using multi modal MR images is essential for the evaluation of the disease progression in order to improve disease diagnosis and treatment planning. We present a new fully automated method for high graded brain tumor segmentation combining sparse autoencoder and multimodal Fuzzy C-means clustering. The approach utilizes multimodal MRI contrast: T1, T2, FLAIR and T1c (contrast-enhanced) for 15 high graded glioma (HGG) subjects. The objective of the proposed study is to segment tumor tissues from HGG including edema and tumor core within edema. The segmentation was performed on the training data of the multimodal brain tumor image segmentation benchmark 2015. Sparse autoencoder, which is an unsupervised learning algorithm, was used to automatically learn features from unlabeled dataset of tumor in order to segment edema. Followed by edema segmentation, tumor core was segmented from edema using multimodal FCM clustering. Evaluating the performance of the segmentation results with the ground truth yields high dice score (DS) of 0.9866±0.01 and 0.9843±0.01 for edema and tumor core respectively and high Jaccard similarity (JS) of 0.9738±0.02 and 0.9692±0.02 for edema and tumor core respectively; showed high accuracy in segmenting the complex tumor structures from multi-contrast MR scans of HGG patients. We also compared our methodology in terms of segmentation efficiency with some recent techniques reported in proceedings of MICCAI-BRATS challenge 2015.

Magnetic resonance (MR) images (MRI) are routinely acquired with high in-plane resolution and lower through-plane resolution. Improving the resolution of such data can be achieved through post-processing techniques knows as super-resolution (SR), with various frameworks in existence. Many of these approaches rely on external databases from which SR methods infer relationships between low and high resolution data. The concept of self super-resolution (SSR) has been previously reported, wherein there is no external training data with the method only relying on the acquired image. The approach involves extracting image patches from the acquired image constructing new images based on regression and combining the new images by Fourier Burst Accumulation. In this work, we present four improvements to our previously reported SSR approach. We demonstrate these improvements have a significant effect on improving image quality and the measured resolution.

The Corpus Callosum (CC) is the largest white matter structure in the brain and subject of many relevant studies. In order to properly analyze this structure in 2D studies, the midsagittal plane (MSP) determination of the CC is required. Usually, this computation is done on structural MR images and transformed to diffusion space when necessary. Furthermore, most existing methods take into account the whole brain structure instead of only the object of study. Differently, our work proposes a plane computation based on the structure of interest, directly on Diffusion Tensor Images (DTI), through the DTI-based divergence map.1 Since our plane is computed in the diffusion domain, the method explores the high organization of the fibers in the CC to establish a reference system that can be used to perform 2D CC studies, while most existing MSP computation algorithms are based on structural characteristics of the brain, such as shape symmetry and inter-hemispheric fissure location. Experiments showed that the proposed method is reliable regarding repeatability and parameters choices. Results also indicate that the callosal fiber convergence plane (CFCP) found by our method is similar to MSP in most subjects. Nevertheless, when the CC is not well aligned with the brain intercommissural fissure, CFCP and MSP presented significant differences.

Cerebrospinal fluid (CSF) is known to be reabsorbed by the lymphatic vessels and drain into the lymph nodes (LNs) through peripheral lymphatic vessels. In the peripheral lymphatics, the contractile pumping action of lymphangions mediates lymph drainage; yet it is unknown whether lymphatic vessels draining cranial and spinal CSF show similar function. Herein, we used non-invasive near-infrared fluorescence imaging (NIRFI) to image (i) indocyanine green (ICG) distribution along the neuraxis and (ii) routes of ICG-laden CSF outflow into the lymphatics following intrathecal lumbar administration. For intrathecal injection, a 31-gauge needle was inserted between L5 and L6 vertebrae and a tail flick response was referenced as indication of correct position of the needle into the intradural space. Imaging agents were injected immediately after a tail flick. A volume of 10 to 30μl of 645μΜ of ICG was injected. We demonstrate lymphatic contractile function in peripheral lymphatics draining from the nasal lymphatics to the mandibular LNs. In addition, we observed afferent sciatic lymphatic vessels, which also show contractile activity and transport spinal CSF into the sciatic LNs. NIRFI could be used as a tool to probe CSF pathology including neurological disorders by imaging CSF outflow dynamics to lymphatics.

Corpus Callosum (CC) is the largest white matter structure and it plays a crucial role in clinical and research studies due to its shape and volume correlation to subject’s characteristics and neurodegenerative diseases. CC segmentation and parcellation are an important step for any MRI-based clinical and research study. There is only a few automatic CC parcellation methods proposed in the literature and, since it is not trivial to build a ground truth, most methods are validated qualitatively. We present a quantitative analysis of different state of art CC parcellation methods aiming to compare their results on a common dataset. Our findings show a significant difference among the same CC parcels, but using different CC parcellation methods, and its impact on the diffusion properties.

Neo-adjuvant chemotherapy (NAC) is the treatment of choice in patients with locally advanced breast cancer to reduce tumor burden, and potentially enable breast conservation. Response to treatment is assessed by histopathology from surgical specimen, a pathological complete response (pCR), or a minimal residual disease are associated with an improved disease-free, and overall survival. Early identification of non-responders is crucial as these patients might require different, or more aggressive treatment. Multi-parametric magnetic resonance imaging (mpMRI) using different morphological and functional MRI parameters such as T2-weighted, dynamic contrast-enhanced (DCE) MRI, and diffusion weighted imaging (DWI) has emerged as the method of choice for the early response assessments to NAC. Although, mpMRI is superior to conventional mammography for predicting treatment response, and evaluating residual disease, yet there is still room for improvement. In the past decade, the field of medical imaging analysis has grown exponentially, with an increased numbers of pattern recognition tools, and an increase in data sizes. These advances have heralded the field of radiomics. Radiomics allows the high-throughput extraction of the quantitative features that result in the conversion of images into mineable data, and the subsequent analysis of the data for an improved decision support with response monitoring during NAC being no exception. In this paper, we determine the importance and ranking of the extracted parameters from mpMRI using T2-weighted, DCE, and DWI for prediction of pCR and patient outcomes with respect to metastases and disease-specific death.

Existing methods for automated breast ultrasound lesions detection and recognition tend to be based on multi-stage processing, such as preprocessing, filtering/denoising, segmentation and classification. The performance of these processes is dependent on the prior stages. To improve the current state of the art, we have proposed an end-to-end breast ultrasound lesions detection and recognition using a deep learning approach. We implemented a popular semantic segmentation framework, i.e. Fully Convolutional Network (FCN-AlexNet) for our experiment. To overcome data deficiency, we used a pre-trained model based on ImageNet and transfer learning. We validated our results on two datasets, which consist of a total of 113 malignant and 356 benign lesions. We assessed the performance of the model using the following split: 70% for training data, 10% for validation data, and 20% testing data. The results show that our proposed method performed better on benign lesions, with a Dice score of 0.6879, when compared to the malignant lesions with a Dice score of 0.5525. When considering the number of images with Dice score > 0.5, 79% of the benign lesions were successfully segmented and correctly recognised, while 65% of the malignant lesions were successfully segmented and correctly recognised. This paper provides the first end-to-end solution for breast ultrasound lesion recognition. The future challenges for the proposed approaches are to obtain additional datasets and customize the deep learning framework to improve the accuracy of this method.

Introduction: Prostate cancer is the second leading cause of cancer death in men. Biopsy serves as the primary tool for cancer diagnoses in these men. However, false-negative diagnosis following biopsy can be as high as 30% and even when detected via biopsy it can be difficult to accurately grade the cancer. Electrical properties of prostate cancer have been reported to be significantly different than benign prostate. We hypothesize that a custom tetrapolar-based electrical impedance sensing biopsy (EIS-Bx) needle will be able to detect electrical properties of surrounding tissue and provide a "4 bit" image for guidance to potential cancer locations. Methods: A custom EIS-Bx device was designed using four goldplated electrode traces on a standard biopsy needle. A novel small form-factor impedance analyzer was designed to interface with the EIS-Bx needle. The EIS-Bx device was submerged in a saline bath while a high contrast inclusion was rotated in 45-degree increments around the needle. At each location, the impedance of 4 electrode configurations was recorded at 7 frequencies (ranging from 1kHz to 100kHz). The impedances of each quadrant were compared with the inclusion location to examine spatial differentiation. Results: Bipolar measurements clearly detected impedance changes correlated to inclusion presence across frequencies. These results validate the hypothesis of potential "4-bit" imaging for cancer detection and diagnostic guidance. Conclusion: Initial experiments successfully demonstrate spatial sensitivity to a moving inclusion using the EIS-Bx device. Future work will investigate the ability to differentiate cancer from benign tissue ex-vivo with quadrant specific resolution and to display this as a real-time map of prostate pathology.

We created a cancer imaging and therapy platform (CITP) consisting of software and multi-spectral cryo-imaging to support innovations in preclinical cancer research. Cryo-imaging repeatedly sections and tiles microscope images of the tissue block face, providing anatomical episcopic color and molecular fluorescence, enabling 3D microscopic imaging of the entire mouse with single metastatic cell sensitivity. Our platform allows tumor molecular imaging validation with MRI and cryo images registration, GFP metastatic tumor segmentation and quantitative analysis, all of which are important processes in the CITP visualization/analysis pipeline. Our standard approach to register MRI to the cryo color volume involves preprocess Æ affine Æ B-spline non-rigid 3D mutual information registration. We further developed modified mask registration to allow improved registration quality within the created 3D cuboid mask on the organ of interest. In 3 mice kidneys, standard and mask registration yields Dice index of 84% ± 2% and 90% ± 2%, respectively. To segment big metastases in GFP, we use marker based watershed with intensity thresholding. For small metastases, we apply Laplacian of Gaussian filtering to get candidate metastases and use morphological features and support vector machine to classify the candidates. In a test mouse, sensitivity/specificity for metastases detection was 94.1%/99.82% as compared with manual segmentation of 202 metastases. Quantitative analysis of molecular MR imaging agent CREKA-Gd using Rose SNR in the lung of a test mouse showed that all micro-metastases ≥ 0.25 mm² were detectable with Rose SNR ≥ 4 and around 36% of micro-metastases < 0.25 mm² were detectable.

The architecture of intra-tumoral vascular network continuously evolves with tumor progression. Non-invasive methods that facilitate 3D in vivo interrogation of tumor vascular architecture could improve understanding of tumor progression and metastasis. In this work, we studied evolving tumor vasculature using high-resolution CT images and a blood-pool, nanoparticle iodinated contrast agent. In vivo studies were performed in a transgenic mouse model of neuroblastoma that exhibit spontaneous bilateral tumors in the adrenals. Animals were divided into three groups based on tumor age: early-age tumor, intermediate-age tumor, old-age tumor. Tumor progression was monitored using T2-weighted MRI. Contrast-enhanced CT imaging was performed at two points: the first imaging session (leak map) was performed 4 days after administration of the nanoparticle agent to interrogate changes in tumor vascular permeability. Immediately thereafter, a second dose of contrast agent was administered and CT imaging was performed within 1 hour to capture high-resolution angiograms of tumor vasculature. CT angiograms demonstrated the highly-vascularized nature of these tumors. Old-age tumors exhibited a higher fractional volume of avascular regions and an increased number of large superficial blood vessels on tumor periphery. Old-age tumors also demonstrated the presence of intra-vessel tumor thrombus and the invasion of tumor into the inferior vena cava. Leak maps images demonstrated signal enhancement throughout the tumor in early-age tumors, including the core region, suggestive of the presence of highly permeable blood vessels through the tumor volume. Old-age tumors exhibited relatively lower signal enhancement, indicative of a less 'leaky' tumor vascular network compared to early and intermediate-stage tumors.

In this study, we proposed and validated a novel and accurate automatic approach for ulcer area extraction from ocular staining images. We first segmented the corneal surface area with the help of four pre-defined key landmarks by modeling the corneal surface shape as an ellipse. Then the ulcer area was identified within the cornea by employing a combination of techniques: 1) iterative k-means based clustering to extract areas with similar color information; 2) morphological operations to polish results from the previous step, with the parameters employed in the morphological operators determined automatically via linear regression analysis; 3) region growing to select the true ulcer area among a number of separated areas. To validate this automatic approach, we compared its results with those from manual delineations using the Dice Overlap Score (DSC) and the automatic-versus-manual correlation in terms of the ulcer area size based on 48 ocular staining images with corneal ulcers. The automatic results showed strong and statistically significant positive correlations with the manual ones in terms of both the cornea size and the ulcer area size (cornea: PCC=0.842, p-value=6:890 x 10^-14; corneal ulcer area: PCC=0.969, p-value=1:119 x 10^-29). For cornea, the DSC between the proposed automatic results and the manual ones is on average 0.989, whereas the average DSC for the ulcer segmentation is 0.879. This suggests a high overlap between the automatic and the manual results for both the cornea and the corneal ulcer area. We also compared the proposed method with a classic segmentation approach (the active contour). Our results revealed a superior performance of the proposed automatic approach in corneal ulcer area identification relative to the active contour (0.879 versus 0.639 in terms of DSC).

Recent advances in CT hardware have renewed interest in the development of contrast agents for molecular CT imaging. Nanoparticle platforms are attractive for CT imaging agent development due to their ability to carry a high payload of imaging moieties, thereby facilitating signal amplification at target site, and ease of surface modification to enable selective in vivo targeting against cells/molecules of interest. In this work, we performed investigations for optimizing an iodine-based liposomal nanoparticle platform for molecular CT imaging applications. Since signal intensity is directly proportional to the imaging moiety concentration, optimization studies were performed to rationally design an iodinated nanoparticle construct with maximal iodine carrying capacity. The effect of particle size, liposomal bilayer composition, iodine moiety and starting iodine concentration were systematically investigated. The in vitro stability of the optimal formulation was evaluated using plasma assay and the in vivo stability was tested by performing longitudinal micro-CT imaging in live animals. Simulations were performed to study the effects of iodine per nanoparticle and iodine contrast sensitivity on detectability of nanoparticles per image voxel. In vitro optimization studies demonstrated that particle size, type of iodine moiety and starting iodine concentration strongly influenced the iodine loading per nanoparticle. A nanoparticle composition was identified that demonstrated highest iodine loading capacity (∼ 8 million iodine atoms per particle). Micro-CT imaging demonstrated in vivo stability of the high-iodine containing nanoparticle construct. Simulation studies demonstrated a non-linear effect of iodine contrast sensitivity and image voxel size on the limit of nanoparticle detectability.

Magnetic Particle Imaging (MPI) is a highly sensitive imaging modality, which allows the visualization of magnetic tracer materials with a temporal resolution of more than 40 volumes per second. In MPI the size of the field of view scales with the strength of the applied magnetic fields. In clinical applications this strength is limited by peripheral nerve stimulation and specific absorption rates. Therefore, the size of the field of view is usually no larger than a few cubic centimeters. To bypass this limitation additional focus fields and/or a external object movements can be applied. In this work we investigate the later approach, where an object is moved through the scanner bore one step at a time, while the MPI scanner continuously acquires data from its static field of view. Using 3D phantom and 3D+t in-vivo data it is shown that the data can be jointly reconstructed after reordering the data with respect to the stepwise object shifts and heart beat phases.

Cone beam X-ray luminescence computed tomography (CB-XLCT) has recently been proposed as a new imaging modality for biological imaging application. Compared with other XLCT systems such as pencil beam XLCT and narrow beam XLCT, CB-XLCT can achieve fast imaging, where the speed is essential to small animal in vivo imaging studies. However, due to the high degree of light scattering in biological tissues, the CB-XLCT reconstruction is an ill-posed problem, which can result in poor image quality such as low spatial resolution. As a hybrid CT/optical imaging technique, the image quality is conjected to be improved substantially with the structural guidance from the anatomical images of the CT component. For that purpose, in this paper, a direct prior regularization method is proposed by introducing anatomical information directly into the CB-XLCT reconstruction. The primary advantage of the proposed method is that it does not require segmentation of targets in the anatomical images. Phantom experiments with different edge-to-edge distance (EED) were performed to realize the proposed approach's feasibility. Phantom experiments results indicate that the proposed direct regularization method can separate two luminescent targets with an EED of 0 mm. Compared with no-prior reconstruction methods such as ART and adaptive Tikhonov methods, the proposed method can significantly improve the imaging resolution of CB-XLCT.

Purpose: In-vivo evaluation of bone microarchitecture remains challenging because of limited resolution of conventional orthopaedic imaging modalities. We investigate the performance of flat-panel detector extremity Cone-Beam CT (CBCT) in quantitative analysis of trabecular bone. To enable accurate morphometry of fine trabecular bone architecture, advanced CBCT pre-processing and segmentation algorithms are developed.

Methods: The study involved 35 transilliac bone biopsy samples imaged on extremity CBCT (voxel size 75 μm, imaging dose ~13 mGy) and gold standard μCT (voxel size 7.67 μm). CBCT image segmentation was performed using (i) global Otsu’s thresholding, (ii) Bernsen’s local thresholding, (iii) Bernsen’s local thresholding with additional histogram-based global pre-thresholding, and (iv) the same as (iii) but combined with contrast enhancement using a Laplacian Pyramid. Correlations between extremity CBCT with the different segmentation algorithms and gold standard μCT were investigated for measurements of Bone Volume over Total Volume (BV/TV), Trabecular Thickness (Tb.Th), Trabecular Spacing (Tb.Sp), and Trabecular Number (Tb.N).

Results: The combination of local thresholding with global pre-thresholding and Laplacian contrast enhancement outperformed other CBCT segmentation methods. Using this optimal segmentation scheme, strong correlation between extremity CBCT and μCT was achieved, with Pearson coefficients of 0.93 for BV/TV, 0.89 for Tb.Th, 0.91 for Tb.Sp, and 0.88 for Tb.N (all results statistically significant). Compared to a simple global CBCT segmentation using Otsu’s algorithm, the advanced segmentation method achieved ~20% improvement in the correlation coefficient for Tb.Th and ~50% improvement for Tb.Sp.

Conclusions: Extremity CBCT combined with advanced image pre-processing and segmentation achieves high correlation with gold standard μCT in measurements of trabecular microstructure. This motivates ongoing development of clinical applications of extremity CBCT in in-vivo evaluation of bone health e.g. in early osteoarthritis and osteoporosis.

Studies show that cracked teeth are the third most common cause for tooth loss in industrialized countries. If detected early and accurately, patients can retain their teeth for a longer time. Most cracks are not detected early because of the discontinuous symptoms and lack of good diagnostic tools. Currently used imaging modalities like Cone Beam Computed Tomography (CBCT) and intraoral radiography often have low sensitivity and do not show cracks clearly. This paper introduces a novel method that can detect, quantify, and localize cracks automatically in high resolution CBCT (hr-CBCT) scans of teeth using steerable wavelets and learning methods. These initial results were created using hr-CBCT scans of a set of healthy teeth and of teeth with simulated longitudinal cracks. The cracks were simulated using multiple orientations. The crack detection was trained on the most significant wavelet coefficients at each scale using a bagged classifier of Support Vector Machines. Our results show high discriminative specificity and sensitivity of this method. The framework aims to be automatic, reproducible, and open-source. Future work will focus on the clinical validation of the proposed techniques on different types of cracks ex-vivo. We believe that this work will ultimately lead to improved tracking and detection of cracks allowing for longer lasting healthy teeth.

Osteoporosis, associated with reduced bone mineral density and structural degeneration, greatly increases the risk of fragility fracture. A major challenge of volumetric bone imaging of the hip is the selection of regions of interest (ROIs) for computation of regional bone measurements. Here, we develop an MRI-based active shape model (ASM) of the human proximal femur used to automatically generate ROIs. Major challenges in developing the ASM of a complex three-dimensional (3-D) shape lie in determining a large number of anatomically consistent landmarks for a set of training shapes. In this paper, we develop a new method of generating the proximal femur ASM, where two types of landmarks, namely fiducial and secondary landmarks, are used. The method of computing the MRI-based proximal femur ASM consists of—(1) segmentation of the proximal femur bone volume, (2) smoothing the bone surface, (3) drawing fiducial landmark lines on training shapes, (4) drawing secondary landmarks on a reference shape, (5) landmark mesh generation on the reference shape using both fiducial and secondary landmarks, (6) generation of secondary landmarks on other training shapes using the correspondence of fiducial landmarks and an elastic deformation of the landmark mesh, (7) computation of the active shape model. An MRI-based shape model of the human proximalfemur has been developed using hip MR scans of 45 post-menopausal women. The results of secondary landmark generation were visually satisfactory and no topology violation or notable geometric distortion artifacts were observed. Performance of the method was examined in terms of shape representation errors in a leave-one-out test. The mean and standard deviation of leave-one-out shape representation errors were 2.27 and 0.61 voxels respectively. The experimental results suggest that the framework of fiducial and secondary landmark allows reliable computation statistical shape models for complex 3-D anatomic structures.

Chronic kidney disease (CKD) is associated with gradual bone loss that occurs from the failure of the kidneys to regulate bone mineralization. Degradation of bone structure can be quantified with the usage of Micro-CT. The current methods of quantitative imaging typically use a single region of interest (ROI) that segments the whole trabecular region and obtain bone parameters, which usually are not homogenous across such a large ROI. Here we introduce a novel method of quantifying bone parameters that can be used to determine overall bone health. This method analyzes sequential regions on the trabecular bone with multiple small ROIs and evaluates the gradients of bone parameters across these ROIs. Two C57Bl/6J mice femur groups were prepared: a control and CKD groups. All femurs were scanned with a Micro-CT system using tube voltage of 60 kV and current of 0.667 mA. Femur volumes were reconstructed with the Feldkamp-Davis-Kress algorithm and were imported into MicroView to perform bone analysis. Six different sequential ROIs were selected at different distances from the growth plate (0.5mm increments). The gradients of bone parameters along the ROI distance for the control and CKD group were compared. Significant differences were found between two groups in the gradients of bone volume density (P = 0.0002), connective density (P = 0.0003), trabecular spacing (P = 0.001), and trabecular number (P = 0.01). As a result, our method identified a sharp change in several parameters representing a novel and biologically significant strategy.

We propose a novel method for nonrigid registration of whole coronary artery sequence models with periodic displacement field. 2D nonrigid registration method is proposed that periodic deformed information is applied into correspondence with whole fluoroscopic angiograms. The proposed methodology is divided into two parts: one cycle's nonrigid registration, spreading periodic displacement information into other cycles. In the first part, a nonrigid registration method of one cycle is implemented and used to compensate for any local shape discrepancy. In the second part, periodic displacement field is spreading into images on other cycles in order to align the whole sequence. Experimental evaluation conducted on a set of 9 fluoroscopic angiograms results in a reduced target registration error, which showed the effectiveness of the proposed methodology.

Direct visualization of the ablated region in the left atrium during radiofrequency ablation (RFA) surgery for treating atrial fibrillation (AF) can improve therapy success rates. Our visualization approach is auto-fluorescence hyperspectral imaging (aHSI), which constructs each hypercube containing 31 auto-fluorescence images of the tissue. We wish to use the spectral information to characterize ablated lesions as being successful or not. In this paper, we reshaped one hypercube to a 2D matrix. Each row (sample) in the matrix represents a pixel in the spatial dimension, and the matrix has 31 columns corresponding to 31 spectral features. Then, we applied k-means clustering to detect ablated regions without a priori knowledge about the lesion. We introduced an accuracy index to evaluate the results of k-means by comparing with the reference truth images quantitatively. To speed-up the detection process, we implemented a grouping procedure to decrease the number of features. The 31 features were divided into four contiguous disjoint groups. In each group, the summation of values yielded a new feature. By the same evaluation method, we found the best 4-feature groups to adequately detect the lesions from all possible combinations. The average accuracy for detection by k-means (k=10) using 31 features was about 74% of reference truth images. And, for using the best grouped 4 features, it was about 95% of that using 31 features. The time cost of 4-feature clustering is about 41% of the 31-feature clustering. We expect that the reduction of time for both acquisition and processing will make possible intraoperative real-time display of ablation status.

The fat that accumulates between the myocardium and the visceral pericardium is called epicardial adipose tissue (EAT). When volume is increased, the EAT can secrete chemicals that influence the development of coronary disease. Volumetric assessment of magnetic resonance imaging (MRI) can quantify EAT, but volume alone gives no information about its distribution across the myocardial surface. In this study, a three-dimensional (3D) modeling technique is developed and used to quantify the distribution of the EAT across the surface of the heart. Dixon MRI scans, which produce a registered 3D set of fat-only and water-only images, were acquired in 11 subjects for a study on exercise intervention. A previously developed segmentation algorithm was used to identify the heart and EAT in six of the scans. Contours were extracted from the labeled images and imported into NX 10, where 3D models of both surfaces were created. Procrustes analysis was used to register the heart models and create an average heart surface. An iterative closest point algorithm was used to register and align the EAT models for calculation of EAT thickness. Rays were cast in directions specified by a spherical parameterization of elevation and azimuthal angles, and intersections of the ray with the EAT surface were used to calculate the thickness at each location. The thickness maps were averaged and then “painted” onto the average heart model, creating a single, integrated model representing the average EAT thickness across the surface of the myocardium.

A non-invasive and accurate modality that can continuously monitor stroke volume (SV) for extended periods of time is desired to allow for more proactive care of an increasing population of patients living with heart failure. Electrical impedance tomography (EIT) has been proposed as a method for accurate, non-invasive, continuous, and long-term SV monitoring. While cardiac EIT has been explored, clinical translation has yet to occur and a standardized method for evaluation and comparison of cardiac EIT images is desired. This work explores an automated process for segmenting and extracting features from the images that allow for evaluation and comparison. A simulation study was conducted using the 4D XCAT model to evaluate the proposed method’s ability to automatically segment and extract features from images reconstructed at various phases of the cardiac cycle. The same procedure was then applied to EIT reconstructions on data collected from five healthy volunteers. The automated segmentation is able to accurately capture the heart region-of-interest (ROI) in various images and extract features, which allows comparison of desired signals across reconstructions. ROI mean conductivity, ROI area, sum of conductivities within the ROI, and ROI maximum conductivity were chosen as promising features from the simulation study, with R² values of 0.61, 0.73, 0.75, and 0.66 for a single heart-cycle, and minimum SV distinguishability of 25.54, 12.16, 12.16, and 17.22 ml. In experimental data, the area feature showed the least variation across individual reconstructions while the sum feature showed the highest variation.

Segmentation of the aorta from CT and MR data is important in order to quantitatively assess diseases of the aorta including aortic dissection and distention of aortic aneurysm, among others. In this paper, we propose a segmentation method to extract exact the 3D boundary of the aorta via graph-cuts segmentation. The graph-cuts technique is able to avoid local minima with global optimization and can be applied to 3D and higher dimension with fast computation. We performed 3D segmentation using this method for five CT data sets. The user selects seed points for aorta region as 'object' and surrounding tissues as 'background' on an axial slice of the 3D CT data and the algorithm calculates the cost of n-link (neighborhood-link) and t-link (terminal-link), and computes the minimum cut separating the aorta from the background by applying the max-flow/min-cut algorithm. Results were validated against manually traced aorta boundaries. The mean Dice Similarity Coefficient for the five 3D segmentations was 0.9381. The 3D segmentation took less than five minutes for data sets of size 512×512×244 to 512×512×284.

Vascular structures are important information for education purpose, surgical planning and analysis. Extraction of blood vessels of the organ is a challenging task in the area of medical image processing and it is the first step before obtaining the structure. It is difficult to get accurate vessel segmentation results even with manually labeling by human being. The difficulty of vessels segmentation is the complicated structure of blood vessels and its large variations that make them hard to recognize. In this paper, we present deep artificial neural network architecture to automatically segment the vessels from computed tomography (CT) image. We proposed deep neural network (DNN) architecture for vessel segmentation from a medical CT volume, which consists of multi deep convolution neural networks to extract features from difference planes of CT data. Due to the problem of varies constrains that we cannot control, we add normalization process to make sure our network will well perform on clinical data. To validate effectiveness and efficiency of our proposed method, we conduct experiments on 20 clinical CT volumes. Our network can yield an average dice coefficient 0.879 on clinical data which better than state-of-the-art methods such as level set, Frangi, and submodular graph cuts.

Fractional Flow Reserve (FFR), the ratio of arterial pressure distal to a coronary lesion to the proximal pressure, is indicative of its hemodynamic significance. This quantity can be determined from invasive measurements made with a catheter, or by using computational methods incorporating models of the the coronary vasculature. One of the inputs needed by a model-based approach for estimating FFR from Computed Tomography Angiography (CTA) images (denoted FFR-CT) is the geometry of the coronary arteries, which requires segmentation of the coronary lumen. Several algorithms have been proposed for coronary lumen segmentation, including the recent application of machine learning techniques. For evaluating these algorithms or for training machine learning algorithms, manual segmentation of the lumen has been considered as ground truth. However, since there is inter-subject variability in manual segmentation, it would be useful to first assess the extent to which this variability affects the predicted FFR values. In the current study, we evaluated the impact of inter-subject variability in manual segmentation on computed FFR, using datasets with three different manual segmentations provided as part of the Rotterdam Coronary Artery Evaluation Framework. FFR was computed using a coronary blood flow model. Our results indicate that variability in manual segmentations on FFR estimates depend on the FFR value. For FFR ≥ 0.97, variability in manual segmentations does not impact FFR estimates, while, for lower FFR values, the variability in manual segmentations leads to significant variability in FFR. The results of this study indicate that researchers should exercise caution when treating manual segmentations as ground truth for estimating FFR from CTA images.

There are several computational methods for estimating myocardial blood flow (MBF) using CT myocardial perfusion imaging (CT-MPI). Previous work has shown that model-based deconvolution methods are more accurate and precise than model-independent methods such as singular value decomposition and max-upslope. However, iterative optimization is computationally expensive and models are sensitive to image noise, thus limiting the utility of low x-ray dose acquisitions. We propose a new processing method, SLICR, which segments the myocardium into super-voxels using a modified simple linear iterative clustering (SLIC) algorithm and quantifies MBF via a robust physiologic model (RPM). We compared SLICR against voxel-wise SVD and voxel-wise model-based deconvolution methods (RPM, single-compartment and Johnson-Wilson). We used image data from a digital CT-MPI phantom to evaluate robustness of processing methods to noise at reduced x-ray dose. We validate SLICR in a porcine model with and without partial occlusion of the LAD coronary artery with known pressure-wire fractional flow reserve. SLICR was ~50 times faster than voxel-wise RPM and other model-based methods while retaining sufficient resolution to show all clinically interesting features (e.g., a flow deficit in the endocardial wall). SLICR showed much better precision and accuracy than the other methods. For example, at simulated MBF=100 mL/min/100g and 100 mAs exposure (50% of nominal dose) in the digital simulator, MBF estimates were 101 ± 12 mL/min/100g, 160 ± 54 mL/min/100g, and 122 ± 99 mL/min/100g for SLICR, SVD, and Johnson-Wilson, respectively. SLICR even gave excellent results (103 ± 23 ml/min/100g) at 50 mAs, corresponding to 25% nominal dose.

Pulmonary function is generally evaluated based on the overall capacity of both lungs; this evaluation is performed by a pulmonary functional test using a spirometer. Diaphragm movement has a direct association with pulmonary function. Therefore, evaluation of diaphragm motion is also helpful for estimating pulmonary function in the lung unit. The purpose of this study was to investigate the utility of dynamic analysis of the diaphragm using dynamic flat-panel detector (FPD) imaging for pulmonary function assessment in the lung unit. Sequential chest radiographs of four pigs (body weight approximately 20−30 kg) were obtained using a dynamic FPD system under respiratory control with a ventilator (100, 200, 300, 400, and 500 mL). Diaphragm excursion was measured and then analyzed the correlation with inspired volume. We also created porcine models of atelectasis by a catheter procedure and investigated whether lungs affected by atelectasis could be detected as reduced diaphragm excursion. To facilitate visual evaluation, temporal cross-sectional images were created, with the x-axis representing time, using a linear interpolation method. There was a strong correlation between inspired volume and diaphragm excursion (r = 0.96). In porcine models of atelectasis, diaphragm movement of an affected lung was restricted and was reduced, on average, to 44% of that in unaffected lungs. Reduction in diaphragm movement was also observed in the temporal cross-sectional images. Dynamic FPD imaging allows for relative pulmonary function assessment based on diaphragm movement, and unilateral abnormalities could be detected as reduced diaphragm excursion, even with a normal inspired volume.

Quantitative analyses of three-dimensional (3D) micro structures in human lungs can provide detailed information to elucidate pulmonary disease progress. The observation of tissue specimens in two-dimensional (2D) slice of lungs has a limited understanding of the whole of 3D lung peripheral structures. A 3D analysis of the human pulmonary acinus is fundamental to understand lung structure. However, we cannot yet thoroughly understand the 3D lung microstructure because of technical limitations. The lung specimens used in this study were fixed, inflated and dried according to the Heitzman method. Each lung specimen was taken in the beam line BL20B2 at SPring8. The present findings may give further insight in the 3D lung microstructure.

Manual identification of capillaries in transverse muscle sections is laborious and time consuming. Although the process of classifying a structure as a capillary is facilitated by (immuno)histochemical staining methods, human judgement is still required in a significant number of cases. This is mainly due to the fact that not all capillaries stain as strongly: they may have an elongated appearance and/or there may be staining artefacts that would lead to a false identification of a capillary. Here we propose two automated methods of capillary detection: a novel image processing approach and an existing machine learning approach that has been previously used to detect nuclei-shaped objects. The robustness of the proposed methods was tested on two sets of differently stained muscle sections. On average, the image processing approach scored a True Positive Rate of 0.817 and a harmonic mean (F1 measure) of 0.804 whilst the machine learning approach scored a True Positive Rate 0.843 and F1 measure of 0.846. Both proposed methods are thus able to mimic most of the manual capillary detection, but further improvements are required for practical applications.

Intravascular optical coherence tomography (IOCT) is a modality that provides sufficient resolution for very accurate visualization of localized cardiovascular conditions, such as coronary artery calcification (CAC). CAC quantification in IOCT images is still performed mostly manually, which is time consuming, considering that each IOCT exam has more than two hundred 2D slices. An automated method for CAC detection in IOCT would add valuable information for clinicians when treating patients with coronary atherosclerosis. In this context, we propose an approach that uses a fully connected neural network (FCNN) for CAC detection in IOCT images using a small training dataset. In our approach, we transform the input image to polar coordinate transformation using as reference the centroid from the lumen segmentation, that restricts the variability in CAC spatial position, which we proved to be beneficial for the CNN training with few training data. We analyzed 51 slices from in-vivo human coronaries and the method achieved 63.6% sensitivity and 99.8% specificity for segmenting CAC. Our results demonstrate that it is possible to successfully detect and segment calcific plaques in IOCT images using FCNNs.

Accurate segmentation of medical images is a key step in medical image processing. As the amount of medical images obtained in diagnostics, clinical studies and treatment planning increases, automatic segmentation algorithms become increasingly more important. Therefore, we plan to develop an automatic segmentation approach for the urinary bladder in computed tomography (CT) images using deep learning. For training such a neural network, a large amount of labeled training data is needed. However, public data sets of medical images with segmented ground truth are scarce. We overcome this problem by generating binary masks of images of the ¹⁸F-FDG enhanced urinary bladder obtained from a multi-modal scanner delivering registered CT and positron emission tomography (PET) image pairs. Since PET images offer good contrast, a simple thresholding algorithm suffices for segmentation. We apply data augmentation to these datasets to increase the amount of available training data. In this contribution, we present algorithms developed with the medical image processing and visualization platform MeVisLab to achieve our goals. With the proposed methods, accurate segmentation masks of the urinary bladder could be generated, and given datasets could be enlarged by a factor of up to 2500.

This paper presents a novel unsupervised segmentation method for 3D medical images. Convolutional neural networks (CNNs) have brought significant advances in image segmentation. However, most of the recent methods rely on supervised learning, which requires large amounts of manually annotated data. Thus, it is challenging for these methods to cope with the growing amount of medical images. This paper proposes a unified approach to unsupervised deep representation learning and clustering for segmentation. Our proposed method consists of two phases. In the first phase, we learn deep feature representations of training patches from a target image using joint unsupervised learning (JULE) that alternately clusters representations generated by a CNN and updates the CNN parameters using cluster labels as supervisory signals. We extend JULE to 3D medical images by utilizing 3D convolutions throughout the CNN architecture. In the second phase, we apply k-means to the deep representations from the trained CNN and then project cluster labels to the target image in order to obtain the fully segmented image. We evaluated our methods on three images of lung cancer specimens scanned with micro-computed tomography (micro-CT). The automatic segmentation of pathological regions in micro-CT could further contribute to the pathological examination process. Hence, we aim to automatically divide each image into the regions of invasive carcinoma, noninvasive carcinoma, and normal tissue. Our experiments show the potential abilities of unsupervised deep representation learning for medical image segmentation.

The penalized weighted least-squares (PWLS) image reconstruction with the widely used edge-preserving nonlocal means (NLM) penalty has shown the potential to significantly improve the image quality for low dose CT (LDCT). Considering the nonlocal weights have significant effects for the smoothness and resolution of the reconstruction, much effort has been made to improve their accuracy. A high quality image of normal dose with less noise and artifacts is sometimes used for the weight’s calculation to further improvement. However, registration should be employed first when misalignment between the low-dose and normal-dose scans cannot be ignored. It will bring an extra work and the effect of registration error on the proposed method are uncertain. The paper aims to propose a new NLM prior model based on normal-dose CT (NDCT) without registration, by predicting nonlocal weights with selecting most similar patch samples from FDCT database. The patch samples are determined by evaluating the similarity between patches from NDCT and the target patch of LDCT. After building up the normal dose based NLM penalty, the PWLS object function is iteratively minimized for reconstruction. Preliminary reconstruction with LDCT data has shown its potential in the structure detail preservation.

Segmentation of organs at risk (OARs) is a key step during the radiation therapy (RT) treatment planning process. Automatic anatomy recognition (AAR) is a recently developed body-wide multiple object segmentation approach, where segmentation is designed as two dichotomous steps: object recognition (or localization) and object delineation. Recognition is the high-level process of determining the whereabouts of an object, and delineation is the meticulous lowlevel process of precisely indicating the space occupied by an object. This study focuses on recognition.

The purpose of this paper is to introduce new features of the AAR-recognition approach (abbreviated as AAR-R from now on) of combining texture and intensity information into the recognition procedure, using the optimal spanning tree to achieve the optimal hierarchy for recognition to minimize recognition errors, and to illustrate recognition performance by using large-scale testing computed tomography (CT) data sets. The data sets pertain to 216 non-serial (planning) and 82 serial (re-planning) studies of head and neck (H&N) cancer patients undergoing radiation therapy, involving a total of ~2600 object samples. Texture property “maximum probability of occurrence” derived from the co-occurrence matrix was determined to be the best property and is utilized in conjunction with intensity properties in AAR-R. An optimal spanning tree is found in the complete graph whose nodes are individual objects, and then the tree is used as the hierarchy in recognition. Texture information combined with intensity can significantly reduce location error for glandrelated objects (parotid and submandibular glands). We also report recognition results by considering image quality, which is a novel concept. AAR-R with new features achieves a location error of less than 4 mm (~1.5 voxels in our studies) for good quality images for both serial and non-serial studies.

Accurate segmentation of organs-at-risk (OAR) is essential for treatment planning of head and neck (HaN) cancers. A desire to shift from manual segmentation to automated processes allows for more efficient treatment planning. However, the technology of automated segmentation is hindered by complex and irregular morphology, poor soft tissue contrast, artifacts from dental fillings, variability of patient's anatomy, and inter-observer variability. In this study, we propose a state-of-the-art automated segmentation of OAR using a multi-output support vector regression (MSVR) machine learning algorithm to address these challenges under various selectable parameters. Shape image features were extracted using the histogram of oriented gradients and ground truth boundaries were obtained from physicians. Automated delineation of the OAR was performed on CT images from 56 subjects consisting of the brain stem, cochleae, esophagus, eye globes, larynx, lenses, lips, mandible, oral cavity, parotid glands, spinal cord, submandibular glands, and thyroid. Testing was done on previously unseen CT images. Model performance was evaluated using the dice similarity coefficient (DSC) and leave-one-subject- out strategy. Segmentation results varied from 66.9% DSC for the left cochlea to 93.8% DSC for the left eye globe. Analysis of the performance of a state-of-the-art algorithm reported in literature compared to MSVR demonstrated similar or superior performance on the segmentation of the OAR listed in this study. The proposed MSVR model accurately and efficiently segmented the OAR using a representative database of 56 HaN CT images. Thus, this model is an effective tool to aid physicians in reducing diagnostic and prognostic time.

For every patient, nowadays, dentists use a software to generate the dental scheme. The dental scheme is basically a diagram representing the whole dentition of the patient. On this diagram, each tooth is represented along with the various operations the patient underwent. The dental scheme for instance shows whether some teeth are missing, or if any treatment was ever performed on the dental roots, it also represents the dental fillings, removable prosthesis, dental crowns or tooth bridges. Filling up the dental scheme may be tedious for dentists, as for every new patient, they would have to carefully make an inventory of every dental care the patient underwent. In this work, we intend to study the feasibility of automatically generating the dental scheme from radiographs. Indeed, we aim to propose an image processing method that would automatically detect missing teeth, as well as any dental care in the dentition, this may save a significant amount of time during the dental consultation. In a first step, our method extracts the relevant portion of the scanner image, i.e. we automatically crop the dentition and thus remove the jaws and chin. The bending of the jaw (dentition curvature) is also estimated, and allows to distinguish the upper and lower jaws. A local minimum/maximum computation coupled with the Hough transform, and a fit with Gaussian Mixture Models helps us to segment the teeth despite strong luminance irregularities due to the imaged spine.

Segmentation of heart substructures in cardiac magnetic resonance (CMR) is an important step in the quantitative assessment of the impact of cardiovascular disease. Manual delineation of these structures, over many patients and multiple time phases, is time consuming and prone to human error and fatigue. In this work we use a deep fully convolutional neural network architecture to automatically segment heart substructures in CMR, achieving state of the art results on a recent benchmark dataset. We further apply our process to a much larger study of CMR subjects, automatically segmenting both left and right ventricular endocardiums (LV, RV) with full thirty-phase time resolution, and LV epicardium (Epi) at end-diastole. We validate our automatically obtained results against manual delineations using Dice overlap and Hausdorff distance, as well as Bland-Altman limits of agreement on the derived blood volumes, ejection fraction, and LV mass. We obtain median Dice overlaps of 0.97, 0.94, and 0.97 on the three structures respectively, and further find small biases and narrow limits of agreement between the two assessments (manual, automatic) of volumes and mass. Our results show promise for the fully automated analysis of the CMR data stream in the near future.

Quantitative imaging of the human spinal cord may provide information for diagnosis and prognosis of multiple pathologies. Diffusion tensor imaging (DTI) of the spinal cord in pediatric populations may offer quantitative indices to assess pathologies such as Chiari I malformation, spinal dysraphic defects including myelomeningocele, and tethered cord. We obtained DTI of the spine in 19 pediatric patients at Vanderbilt Children’s Hospital. 14 patients were identified as normal, and 5 patients were diagnosed with neurological disorders localized to the spinal cord, including Chiari I, tethered cord, and an intra-spinal tumor. The ages ranged 1 to 16 years old (mean age: 7 years, SD: 4 years). Fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD) were calculated from DTI and spinal cord volumes of interest were assessed. Data from control patients were used to define the range of normal values, and patients with diagnosed disorders affecting the cord were compared against this normal range. DTI data showed significant differences (p<0.05) in FA, AD, and RD between the conus medularis and thoracolumbar cord. Patients with tethered cord demonstrated higher FA, MD, and AD in the conus, and higher AD in the in thoracolumbar cord. The Chiari I patient had higher MD and AD in cervical cord. Whereas, the spinal tumor patient demonstrated no significant trends in the cervical cord.

Tourette syndrome (TS) is a childhood-onset neurobehavioral disorder and its pathophysiological mechanism remains elusive. At present, TS-related abnormalities in either structural connectivity (SC) or functional connectivity (FC) have extensively been described, and discrepancies were apparent between the SC and FC studies. However, abnormalities in the SC-FC correlation for early TS children remain poorly understood. In our study, we used probabilistic diffusion tractography and resting-state FC to construct large-scale structural and functional brain networks for 34 drug-naive TS children and 42 healthy children. Graph theoretical approaches were employed to divide the group-averaged FC networks into functional modules. The Pearson correlation between the entries of SC and FC were estimated as SC-FC coupling within whole-brain and each module. Although five common functional modules (including the sensorimotor, default-mode, fronto-parietal, temporo-occipital and subcortical modules) were identified in both groups, we found SC– FC coupling in TS exhibited increased at the whole-brain and functional modular level, especially within sensorimotor and subcortical modules. The increased SC-FC coupling may suggest that TS pathology leads to functional interactions that are more directly related to the underlying SC of the brain and may be indicative of more stringent and less dynamic brain function in TS children. Together, our study demonstrated that altered whole-brain and module-dependent SC-FC couplings may underlie abnormal brain function in TS, and highlighted the potential for using multimodal neuroimaging biomarkers for TS diagnosis as well as understanding the pathophysiologic mechanisms of TS.

Imaging connectomics emerged as an important field in modern neuroimaging to represent the interaction of structural and functional brain areas. Static graph networks are the mathematical structure to capture these interactions modeled by Pearson correlations between the representative area signals. Dynamical functional resting state networks seen in most fMRI experiments can not be represented by the classic correlation graph network. The changes in brain connectivity observed in many neuro-degenerative diseases in longitudinal data series suggest that more sophisticated graph networks to capture the dynamical properties of the brain networks are required. Furthermore, certain brain areas seem to act as ”disease epicenters” being responsible for the spread of neuro-degenerative diseases. To mathematically describe these aspects, we propose a novel framework based on pinning controllability applied to dynamic graphs and seek to determine the changes in the ”driver nodes” during the course of the disease. In contrast to other current research in pinning controllability, we aim to identify the best driver nodes describing disease evolution with respect to connectivity changes and location of the best driver nodes in functional ¹⁸F-Fluorodeoxyglucose Positron Emission Tomography (¹⁸FDG-PET) and structural Magnetic Resonance Imaging (MRI) connectivity graphs in healthy controls (CN), and patients with mild cognitive impairment (MCI), and Alzheimer’s disease (AD). We present the theoretical framework for determining the best driver nodes in connectivity graphs and their relation to disease evolution in dementia. We revolutionize the current graph analysis in brain networks and apply the concept of dynamic graph theory in connection with pinning controllability to reveal differences in the location of ”disease epicenters” that play an important role in the temporal evolution of dementia. The described research will constitute a leap in biomedical research related to novel disease prediction trajectories and precision dementia therapies.

Multivariate classification techniques have been widely applied to decode brain states using functional magnetic resonance imaging (fMRI). Because the number of labeled samples is limited by the financial and safety consideration during fMRI data acquirement, it is not easy to train a robust classifier for fMRI data. Recently, semi-supervised learning has been proposed to train the classifier using both labeled training data and unlabeled data. Moreover, sparse representation based classification (SRC) has seldom been applied to fMRI data, although it exhibits a state-of-the-art classification performance in image processing. In this study, we proposed semi-supervised SRC with random sample subset ensemble strategy (semiSRC-RSSE) that used the average of class-specific coefficients as the SRC classification criterion and dynamically update the training dataset using the random sample subset ensemble method to measure the confidence of the prediction of each test sample. The results of the simulated and real fMRI experiments showed that semiSRC-RSSE method largely improved the classification accuracy of SRC and had better performance than support vector machine (SVM) and semi-supervised SVM with the random sample subset ensemble strategy (semiSVM-RSSE).

Although fMRI technique can be used to scan the whole brain to investigate the brain functional activities in children by using the blood oxygenation level dependent (BOLD) MRI signal, patients often require sedation, which prevents investigation of brain function during the actual disordered movements. Functional near infrared spectroscopy and diffuse optical tomography (fNIRS-DOT) allows a similar assessment to the BOLD MRI signal. However, in a realistic situation there will be a skull, blood vessels, cerebrospinal fluid and a significant variation of distance of the deep hemovascular target from the tissue surface. Transmitted diffuse near-infrared (NIR) light is significantly scattered and attenuated, making it difficult to be detected with sufficient signal-to-noise ratio (SNR) using conventional detectors such as avalanche photodiodes or photomultiplier tubes. In this study, we present an fNIRS-DOT imaging system by use of military based intensification with charge-coupled device (CCD) technology to acquire the transmitted weak diffuse NIR signals with high sensitivity that allows whole brain imaging and the device can be mounted in a comfortable cap on an awake child. The millimeter level spatial resolution of transillumination tomographic reconstructions was demonstrated in studies of phantoms that approximate the size of a child’s brain. In addition, we have been able to obtain preliminary transcranial proof-of-concept data with the reasonable SNR in a 7 month old healthy baby. Further tomographic studies will allow the assessment of the brain network dysfunction in awake children suffering from brain diseases such as brain tumor.

Medical imaging is an integral part of today's world. Many diseases can be diagnosed with an examination of the inner structure of the body. Electrical impedance tomography (EIT) is an imaging method which is used in the medical field. In addition to the very widespread use of lung diagnostics, the EIT also finds application for cancer care in the female breast area. A non-existent application is the image reconstruction of the human extremities, especially arms and legs. In the future a fast imaging using a mobile EIT system can help to create a first diagnosis according to diseased tissue. Thus it is possible to decide on the spot whether a further treatment in a hospital is necessary. However, there is as yet no mobile EIT system that allows such a diagnosis on site. Current EIT systems are not suitable for mobile application which is a major obstacle for exploring the techniques capabilities. For this application, an EIT system has been designed that is mobile and allows fast image analysis. It is powered by a rechargeable battery and offers a wireless interface to connect to a host. The practical evaluation is carried out with the determination of the measuring functionality on a phantom. In the next step first measurements are presented on the human body.

Background: Cell permeable peptides (CPP) are a new class of carrier molecule to deliver biomolecules, radio-nucleotide and drugs that is gaining momentum. CPP are capable of entering into the cells by breaking the resistance of the membrane barrier and thus can be used universally in many cell types, which renders it an efficient carrier for both in-vitro and in-vivo use.

Methods: L-Maurocalcine (L-MCa), a peptide derived from scorpion venom was radiolabeled with ¹²⁵I using the lactoperoxidase method. We achieved a specific activity of 45Mbq/nmole. In vitro studies with ¹²⁵I-L-MCa in DAOY cells (human medulloblastoma) were studied in order to analyze the uptake of the peptide. ¹²⁵I-L-MCa was injected intravenously in mice through tail vein and bio-distribution was studied using single photon emission tomography/computed tomography (SPECT/CT).

Results: The cellular uptake of the ¹²⁵I-L-MCa in DAOY cells was time and dose dependent suggesting that the radiolabeled peptide retains the biological property after radiolabeling. We have observed no loss of cell viability upon uptake of ¹²⁵I-L-MCa, favoring that this peptide has potential for use in in vivo studies. The distribution of the ¹²⁵I-L-MCa in mice revealed its uptake in the liver, kidney and stomach. Interestingly the ¹²⁵I-L-MCa was cleared from the circulation 24h post injection, thus providing another advantage for its use in in vivo studies.

Conclusions: In the present study we have shown the uptake of ¹²⁵I-L-MCa in DAOY cells. Further, the ¹²⁵I-L-MCa when injected in mice localized to the liver, kidney and stomach as revealed by SPECT/CT. Cells labeled with ¹²⁵I-L-MCa can possibly be tracked to their target site.

The growing use of superparamagnetic iron oxide nanoparticles (SPIONs) in early cancer detection technologies has created a demand for physiologically-based pharmacokinetic (PBPK) models that accurately model and predict the biodistribution of SPIONs in the mouse and human model. The objective of this work is to use a Bayesian approach built upon nested-sampling to select a model based on qualitative criteria of the fit of the model and the likelihood function landscape, as well as quantitative criteria of the evidence and maximum likelihood values. Four first-order PBPK compartmental models of ranging complexity are considered. Compartments included in the models comprise of a combination of the plasma, liver, spleen, tumor, and “other” (the remaining body tissue), with parameters including the volume, blood flow rate, and plasma:tissue distribution ratios. The model parameters for each model are evaluated using Bayesian inference, in addition to the respective evidence integrals, maximum log-likelihoods, and Bayes factors. The model containing all compartments and the model containing the plasma, liver, tumor and “other” had the highest log-likelihood and evidence values, indicating both a high goodness-of-fit and a high likelihood of the model given the data. This is similarly reflected in a faithful quality-of-fit and non-flat log-likelihood landscapes. Overall, these findings illustrate the strength of the Bayesian model selection framework in ranking different models to determine the best model that accurately represents the experimental data.

Temporo-mandibular osteo arthritis (TMJ OA) is characterized by progressive cartilage degradation and subchondral bone remodeling. The causes of this pathology remain unclear. Current research efforts are concentrated in finding new biomarkers that will help us understand disease progression and ultimately improve the treatment of the disease. In this work, we present Shape Variation Analyzer (SVA), the goal is to develop a noninvasive technique to provide information about shape changes in TMJ OA. SVA uses neural networks to classify morphological variations of 3D models of the mandibular condyle. The shape features used for training include normal vectors, curvature and distances to average models of the condyles. The selected features are purely geometric and are shown to favor the classification task into 6 groups generated by consensus between two clinician experts. With this new approach, we were able to accurately classify 3D models of condyles. In this paper, we present the methods used and the results obtained with this new tool.

DNA is packaged into chromosomes, which occupy a specific region in the three-dimensional (3D) nuclear space known as the chromosome territories (CTs). The spatial organization (SO) of CTs within the nucleus is non-random and any disruption of this organization leads to undesired changes, such as disease states. Determining how CTs organize in the nucleus can allow us to unravel any changes occurring during aneuploidy (loss or gain of chromosomes), a hallmark of cancer. Here, we describe a 3D modeling approach to allow precise shape estimation and localization of CTs in the nucleus of human embryonic stem cells (hES) undergoing progressive but defined aneuploidy. The hES cell line WA09 acquires an extra copy of chromosome 12 in culture with increasing passages. Both diploid and aneuploid nuclei were analyzed to quantitate the differences in the localization of CTs for chromosome 12 as it transitions from euploidy to aneuploidy. The CTs were detected with chromosome specific DNA probes via multi-color fluorescence in situ hybridization (FISH) in conjunction with confocal microscopy. We then employed spherical harmonic (SPHARM) surface modeling to generate a well-defined 3D surface for both the nuclei and enclosed CTs, thereby allowing precise quantification of their size and shape. The estimated models were compared across multiple cells by aligning the nuclei to a well-defined template followed by determining CT position with respect to a local landmark. Our results present evidence of statistically significant changes in the relative positioning of CTs in trisomy-12 cells when compared to diploid cells from the same population.

While Computerised Tomography (CT) may have been the first clinical tool to study human brains when any suspected abnormality related to the brain occurs, the volumes of CT lesions usually are usually disregarded due to variations among inter-subject measurements. This research responds to this challenge by applying the state of the art deep learning techniques to automatically delineate the boundaries of abnormal features, including tumour, associated edema, head injury, leading to benefiting both patients and clinicians in making timely accurate clinical decisions. The challenge with the application of deep leaning based techniques in medical domain remains that it requires datasets in great abundance, whilst medical data tend to be in small numbers. This work, built on the large field of view of DeepLab convolutional neural network for semantic segmentation, highlights the approaches of both semantics-based and patch-based segmentation to differentiate tumour, lesion and background of the brain. In addition, fusions with a number of other methods to fine tune regional borders are also explored, including conditional random fields (CRF) and multiple scales (MS). With regard to pixel level accuracy, the averaged accuracy rates for segmentation of tumour, lesion and background amount to 82.9%, 85.7%, 85.3% and 81.3% while applying the approaches of DeepLab, DeepLab with MS, DeepLab with MS and CRF, and patch-based pixel-wise classification respectively. In terms of the measurement of intersection over union of two regions, the accuracy rates are of 70.3%, 75.1%, 77.2%, and 63.6% respectively, implying overall DeepLab fused with MS and CRF performs the best.

Adenocarcinomas (ADC) is the major subtype of non-small cell lung cancers. Currently, surgery is used as the main approach for the treatment of the early-stage ADCs. However, different histological subtypes of ADC classified by the IASLC/ATS/ERS system may potentially impact on the surgical management, which subsequently influence the prognosis of the surgery. Thus, preoperative determination of ADC subtypes is essential and highly desirable. Nevertheless, the histological subtypes of ADCs may be either unknown or incompletely determined by biopsy before the surgery.

Alternatively, the histological subtypes of ADCs may be predicted from the pulmonary computed tomographic (CT) images. However, previous studies showed limitations on the prediction results due to the complex composition of ADC subtypes. One possible reason is the radiomic descriptors used to differentiate different subtypes could be very different. The conventional approaches based on the same set of descriptors to distinguish all subtypes are inherently infeasible. Another possible reason is the complex composition of multiple subtypes in a lung nodule may hinder the extraction of effective radiomic descriptors to characterize each subtype. To overcome these challenges, a competing round-robin prediction model was proposed to predict the histological subtypes of ADCs, which was composed of three key ideas, namely, pair-specific radiomic descriptors for differentiation of every pair of subtypes, inter-regional descriptors for characterization of complex composition of subtypes in a nodule, and a multi-level round-robin classifier.

Based on 70 ADCs patients, the proposed model achieved an accuracy of 86.3% in predicting five histological subtypes of adenocarcinomas.

Accurate segmentation of spectral-domain Optical Coherence Tomography (SD-OCT) images helps diagnose retinal pathologies and facilitates the study of their progression/remission. Manual segmentation is clinical-expertise dependent and highly time-consuming. Furthermore, poor image contrast due to high-reflectivity of some retinal layers and the presence of heavy speckle noise, pose severe challenges to the automated segmentation algorithms. The first step towards retinal OCT segmentation therefore, is to create a noise-free image with edge details still preserved, as achieved by image reconstruction on a wavelet-domain preceded by bilateral-filtering. In this context, the current study compares the effects of image denoising using a simple Gaussian-filter to that of wavelet-based denoising, in order to help investigators decide whether an advanced denoising technique is necessary for accurate graph-based intraretinal layer segmentation. A comparative statistical analysis conducted between the mean thicknesses of the six layers segmented by the algorithm and those reported in a previous study, reports non-significant differences for five of the layers (p > 0.05) except for one layer (p = 0.04), when denoised using Gaussian-filter. Non-significant layer thickness differences are seen between both the algorithms for all the six retinal layers (p > 0.05) when bilateral-filtering and wavelet-based denoising is implemented before boundary delineation. However, this minor improvement in accuracy is achieved at an expense of substantial increase in computation time (∼10s when run on a specific CPU) and logical complexity. Therefore, it is debatable if one should opt for advanced denoising techniques over a simple Gaussian-filter when implementing graph-based OCT segmentation algorithms.

In this contribution, the preparation of data for training deep learning networks that are used to segment the lower jawbone in computed tomography (CT) images is proposed. To train a neural network, we had initially only ten CT datasets of the head-neck region with a diverse number of image slices from the clinical routine of a maxillofacial surgery department. In these cases, facial surgeons segmented the lower jawbone in each image slice to generate the ground truth for the segmentation task. Since the number of present images was deemed insufficient to train a deep neural network efficiently, the data was augmented with geometric transformations and added noise. Flipping, rotating and scaling images as well as the addition of various noise types (uniform, Gaussian and salt-and-pepper) were connected within a global macro module under MeVisLab. Our macro module can prepare the data for general deep learning data in an automatic and flexible way. Augmentation methods for segmentation tasks can easily be incorporated.

The recent introduction of next generation spectral optical coherence tomography (OCT) has become increasingly important in the detection and investigation of retinal related diseases. However, unstable eye position of patient makes tracking disease progression over short period difficult. This paper proposed a method to remove the eye position difference for longitudinal retinal OCT data. In the proposed method, pre-processing is first applied to get the projection image. Then, a vessel enhancement filter is applied to detect vessel shadows. Third, SURF algorithm is used to extract the feature points and RANSAC algorithm is used to remove outliers. Finally, transform parameter is estimated and the longitudinal OCT data are registered. Simulation results show that our proposed method is accurate.

Traditional region growing is a semi-automatic segmentation method, which needs manually labeled seeds and is easily leaked to neighbor tissues. In this paper, a new automatic segmentation method based on template matching and improved region growing is proposed. The proposed method consists of three main steps: First, bone is removed from CT images. Second, a multi-scale Gaussian template matching method is used to locate the lung tumors in the three-dimension PET images, in order to eliminate the liver and heart with similar voxel values but with different texture. Third, the seeds are automatically obtained in the process of template matching, in order to make the segmentation process fully automatic. The Euclidean distance measure is added to the criterion of region growth, in order to accelerate the convergence of segmentation and avoid the over-segmentation effectively. The proposed method was tested on 10 PET-CT images from 5 patients with lung tumors, and the average DSC (Dice Similarity Coefficient), TPR (True Positive Ratio) and FPR (False Positive Ratio) were 86.91±5.83%, 85.78±7.34% and 0.033±0.003%, respectively.

Coronary artery disease (CAD) is one of the leading causes of death worldwide. The computed tomography angiography (CTA) is increasingly used to diagnose CAD due to its non-invasive nature and high-resolution three-dimensional (3D) imaging capability of the coronary artery anatomy. CTA allows for identification and grading of stenosis by evaluating the degree of narrowing of the blood-filled coronary artery lumen. Both identification and grading rely on the precise segmentation of the coronary arteries on CTA images. In this paper, a fully automatic segmentation framework is proposed to extract the coronary arteries from the whole cardiac CTA images. The framework adopts a paired multi-scale 3D deep convolutional neural networks (CNNs) to identify which voxels belong to the vessel lumen. Voxels that may belong to coronary artery lumen are recognized by the first CNN in the pair and both artery positives and artery-like negatives are distinguished by the second one. Each CNN is assigned to a different task. They share the same architecture in common but with different weights. In order to combine local and larger contextual information, we adopt a dual pathway architecture that can process the input image simultaneously on multiple scales. The experiments were performed on a CTA dataset from 44 patients. 35 CTA scans are used for training and the rests for testing. The proposed segmentation framework achieved a mean Dice similarity coefficient (DSC) of 0.8649 and mean surface distance (MSD) of 0.5571 with reference to manual annotations. Experimental results show that the proposed framework is capable of performing complete, accurate and robust segmentation of the coronary arteries.

CT is routinely used for radiotherapy planning with organs and regions of interest being segmented for diagnostic evaluation and parameter optimization. For cardiac segmentation, many methods have been proposed for left ventricular segmentation, but few for simultaneous segmentation of the entire heart. In this work, we present a convolutional neural networks (CNN)-based cardiac chamber segmentation method for 3D CT with 5 classes: left ventricle, right ventricle, left atrium, right atrium, and background. We achieved an overall accuracy of 87.2% ± 3.3% and an overall chamber accuracy of 85.6 ± 6.1%. The deep learning based segmentation method may provide an automatic tool for cardiac segmentation on CT images.

An important goal in neuroscience has been to map the surface of the human brain, and many researchers have developed sophisticated methods to parcellate the cortex. However, many of these methods stop short of developing a framework to apply existing cortical maps to new subjects in a consistent fashion. The computationally complex step is often the initial mapping of a large set of brains, and it is inefficient to repeat these processes for every new data sample. In this analysis, we propose the use of a library of training brains to build a statistical model of the parcellated cortical surface and to act as templates for mapping new data. We train classifiers on training data sampled from local neighborhoods on the cortical surface, using features derived from training brain connectivity information, and apply these classifiers to map the surfaces of previously unseen brains. We demonstrate the performance of 3 different classifiers, each trained on 3 different types of training features, to accurately predict the map of new brain surfaces.