Purpose: Existing methods for checking the light field–radiation field congruence on x-ray equipment either do not fully meet the conditions of various quality control standards regarding inherent uncertainty requirements or contain subjective steps, further increasing the uncertainty of the end result. The aim of this work was to develop a method to check the light field–radiation field congruence on all x-ray equipment. The result should have a low uncertainty which is accomplished by eliminating most subjective user steps in the method. A secondary aim was to maintain the same level of usability as of comparable methods but still able to store the result.

Approach: A new device has been developed where the light field and corresponding radiation field are monitored through measurements of the field edge locations (in total: 2  ×  4  edges). The maximum field size location deviation between light field and radiation field in the new method is constrained by the physical limitations of the sensors used in various versions of the prototype: linear image sensors (LISs) of 25 to 29 mm active sensor length. The LISs were sensitized to x-rays by applying a phosphor strip of Gd₂O₂S  :  Tb covering the light sensor input area. Later prototypes of the completed LIS device also have the option of a Bluetooth (100-m range standard) connection, thus increasing the mobility.

Results: The developed device has a special feature of localization a field edge without any prior, subjective, alignment procedure of the user, i.e., the signals produced were processed by software storing the associated field edge profiles, localizing the edges in them, and finally displaying the calculated deviation. The uncertainty in field edge location difference was estimated to be <0.1  mm (k  =  2). The calculated uncertainty is lower than for other, commercially available, methods for light field–radiation field congruence also presented in this work.

Conclusions: A method to check the light field–radiation field congruence of x-ray systems was developed to improve the limitations found in existing methods, such as device detector resolution, subjective operator steps, or the lack of storing results for later analysis. The development work overcame several challenges including mathematically describing real-life edges of light and radiation fields, noise reduction of radiation edges, and mapping/quantification of the rarely observed phenomenon of focal spot wandering. The assessment of the method showed that the listed limitations were overcome, and the aims were accomplished. It is therefore believed that the device can improve the work in quality controls of x-ray systems.

Purpose: To show that adjustment of velocity encoding (VENC) for phase-contrast (PC) flow volume measurements is not necessary in modern MR scanners with effective background velocity offset corrections.

Approach: The independence on VENC was demonstrated theoretically, but also experimentally on dedicated phantoms and on patients with chronic aortic regurgitation (n  =  17) and one healthy volunteer. All PC measurements were performed using a modern MR scanner, where the pre-emphasis circuit but also a subsequent post-processing filter were used for effective correction of background velocity offset errors.

Results: The VENC level strongly affected the velocity noise level in the PC images and, hence, the estimated peak flow velocity. However, neither the regurgitant blood flow volume nor the mean flow velocity displayed any clinically relevant dependency on the VENC level. Also, the background velocity offset was shown to be close to zero (<0.6  cm  /  s) for a VENC range of 150 to 500  cm  /  s, adding no significant errors to the PC flow volume measurement.

Conclusions: Our study shows that reliable PC flow volume measurements are feasible without adjustment of the VENC parameter. Without the need for VENC adjustments, the scan time can be reduced for the benefit of the patient.

Purpose: To assess the physical performance of deep learning image reconstruction (DLIR) compared with those of filtered back projection (FBP) and iterative reconstruction (IR) and to estimate the dose reduction potential of the technique.

Approach: A cylindrical water bath phantom with a diameter of 300 mm including two rods composed of acrylic and soft tissue-equivalent material was scanned using a clinical computed tomography (CT) scanner at four dose levels (CT dose index of 20, 15, 10, and 5 mGy). Phantom images were reconstructed using FBP, DLIR, and IR. The in-plane and z axis task transfer functions (TTFs) and in-plane noise power spectrum (NPS) were measured. The dose reduction potential was estimated by evaluating the system performance function calculated from TTF and NPS. The visibilities of a bar pattern phantom placed in the same water bath phantom were compared.

Results: The use of DLIR resulted in a notable decrease in noise magnitude. The shift in peak NPS frequency was reduced compared with IR. Preservation of in-plane TTF was superior using DLIR than using IR. The estimated dose reduction potentials of DLIR and IR were 39% to 54% and 19% to 29%, respectively. However, the z axis resolution was decreased with DLIR by 6% to 21% compared with FBP. The bar pattern visibilities were approximately consistent with the TTF results in both planes.

Conclusions: The in-plane edge-preserving noise reduction performance of DLIR is superior to that of IR. Moreover, DLIR enables approximately half-dose acquisitions with no deterioration in noise texture in cases that permit some z axis resolution reduction.

Purpose: The purpose of our work is to present a method that utilizes high-quality non-contrast CT (NCCT) images to reduce the noise of CT perfusion (CTP) baseline images to improve the visibility of infarct core in cerebral blood volume (CBV) maps.

Methods: First, a theoretical analysis of the CTP imaging system was performed to demonstrate that for both deconvolution- and non-deconvolution-based CTP systems. The noise of CBV maps is profoundly influenced by the baseline image noise. Consequently, baseline noise reduction is extremely effective in improving the contrast-to-noise ratio (CNR) of ischemic lesions in CBV maps. Second, a method was proposed to fuse the freely available NCCT images with the original CTP baseline images. An optimal weighting scheme was derived such that the noise of the fused baseline image is minimized. Third, the impact of the proposed NCCT-baseline fusion method was investigated using five in vivo canine subjects with different infarct core sizes. NCCT and CTP scans were performed following a clinical stroke CT imaging protocol using a 64-slice MDCT. Two of the subjects also received a diffusion-weighted imaging scan using a 3T-MRI scanner to establish the reference diagnosis for the infarct core.

Results: For all five canine subjects, the proposed method led to lower CBV noise and better conspicuity of the infarct core. Compared with a standard CTP postprocessing method, the proposed method reduced the CBV noise standard deviation by 70  %    ±  24  %   and increased the CNR of infarct core by 23  %    ±  11  %   (p  <  0.01).

Conclusions: By utilizing the high-quality NCCT images to reduce CTP baseline image noise, the quality of CBV maps and the conspicuity of ischemic infarct core can be effectively improved. The proposed method can be readily implemented with minimal interruption to the existing clinical workflow.

Purpose: Hippocampus contouring for radiotherapy planning is performed on MR image data due to poor anatomical visibility on computed tomography (CT) data. Deep learning methods for direct CT hippocampus auto-segmentation exist, but use MR-based training contours. We investigate if these can be replaced by CT-based contours without loss in segmentation performance. This would remove the MR not only from inference but also from training.

Approach: The hippocampus was contoured by medical experts on MR and CT data of 45 patients. Convolutional neural networks (CNNs) for hippocampus segmentation on CT were trained on CT-based or propagated MR-based contours. In both cases, their predictions were evaluated against the MR-based contours considered as the ground truth. Performance was measured using several metrics, including Dice score, surface distances, and contour Dice score. Bayesian dropout was used to estimate model uncertainty.

Results: CNNs trained on propagated MR contours (median Dice 0.67) significantly outperform those trained on CT contours (0.59) and also experts contouring manually on CT (0.59). Differences between the latter two are not significant. Training on MR contours results in lower model uncertainty than training on CT contours. All contouring methods (manual or CNN) on CT perform significantly worse than a CNN segmenting the hippocampus directly on MR (median Dice 0.76). Additional data augmentation by rigid transformations improves the quantitative results but the difference remains significant.

Conclusions: CT-based training contours for CT hippocampus segmentation cannot replace propagated MR-based contours without significant loss in performance. However, if MR-based contours are used, the resulting segmentations outperform experts in contouring the hippocampus on CT.

Purpose: To comprehensively outline the methodology of a fully automated, MRI motion-guided, left-ventricular (LV) chamber quantification algorithm that enhances a similar, existing semi-automated approach. Additionally, to validate the motion-guided technique in comparison to chamber quantification with a vendor tool in post-chemotherapy breast cancer patients susceptible to cardiotoxicity.

Approach: LV deformation data were acquired with the displacement encoding with stimulated echoes (DENSE) sequence on N  =  21 post-chemotherapy female patients and N  =  21 age-matched healthy females. The new chamber quantification algorithm consists of detecting LV boundary motion via a combination of image quantization and DENSE phase-encoded displacements. LV contractility was analyzed via chamber quantification and computations of 3D strains and torsion. For validation, estimates of chamber quantification with the motion-guided algorithm on DENSE and steady-state free precession (SSFP) acquisitions, and similar estimates with an existing vendor tool on DENSE acquisitions were compared via repeated measures analysis. Patient results were compared to healthy subjects for observing abnormalities.

Results: Repeated measures analysis showed similar LV ejection fractions (LVEF), 59  %    ±  6  %  , 58  %    ±  6  %  , and 58  %    ±  6  %  , p  =  0.2, by applying the motion-guided algorithm on DENSE and SSFP and vendor tool on DENSE acquisitions, respectively. Differences found between patients and healthy subjects included enlarged basal diameters (5.0  ±  0.5  cm versus 4.4  ±  0.5  cm, p  <  0.01), torsions (p  <  0.001), and longitudinal strains (p  <  0.001), but not LVEF (p  =  0.1).

Conclusions: Measurement similarities between new and existing tools, and between DENSE and SSFP validated the motion-guided algorithm and differences found between subpopulations demonstrate the ability to detect contractile abnormalities.

Purpose: Convolutional neural network (CNN) methods have been proposed to quantify lesions in medical imaging. Commonly, more than one imaging examination is available for a patient, but the serial information in these images often remains unused. CNN-based methods have the potential to extract valuable information from previously acquired imaging to better quantify lesions on current imaging of the same patient.

Approach: A pretrained CNN can be updated with a patient’s previously acquired imaging: patient-specific fine-tuning (FT). In this work, we studied the improvement in performance of lesion quantification methods on magnetic resonance images after FT compared to a pretrained base CNN. We applied the method to two different approaches: the detection of liver metastases and the segmentation of brain white matter hyperintensities (WMH).

Results: The patient-specific fine-tuned CNN has a better performance than the base CNN. For the liver metastases, the median true positive rate increases from 0.67 to 0.85. For the WMH segmentation, the mean Dice similarity coefficient increases from 0.82 to 0.87.

Conclusions: We showed that patient-specific FT has the potential to improve the lesion quantification performance of general CNNs by exploiting a patient’s previously acquired imaging.

Purpose: Generalizability is an important problem in deep neural networks, especially with variability of data acquisition in clinical magnetic resonance imaging (MRI). Recently, the spatially localized atlas network tiles (SLANT) can effectively segment whole brain, non-contrast T1w MRI with 132 volumetric labels. Transfer learning (TL) is a commonly used domain adaptation tool to update the neural network weights for local factors, yet risks degradation of performance on the original validation/test cohorts.

Approach: We explore TL using unlabeled clinical data to address these concerns in the context of adapting SLANT to scanning protocol variations. We optimize whole-brain segmentation on heterogeneous clinical data by leveraging 480 unlabeled pairs of clinically acquired T1w MRI with and without intravenous contrast. We use labels generated on the pre-contrast image to train on the post-contrast image in a five-fold cross-validation framework. We further validated on a withheld test set of 29 paired scans over a different acquisition domain.

Results: Using TL, we improve reproducibility across imaging pairs measured by the reproducibility Dice coefficient (rDSC) between the pre- and post-contrast image. We showed an increase over the original SLANT algorithm (rDSC 0.82 versus 0.72) and the FreeSurfer v6.0.1 segmentation pipeline (rDSC  =  0.53). We demonstrate the impact of this work decreasing the root-mean-squared error of volumetric estimates of the hippocampus between paired images of the same subject by 67%.

Conclusion: This work demonstrates a pipeline for unlabeled clinical data to translate algorithms optimized for research data to generalize toward heterogeneous clinical acquisitions.

Purpose: Deformable registration problems are conventionally posed in a regularized optimization framework, where balance between fidelity and prescribed regularization usually needs to be tuned for each case. Even so, using a single weight to control regularization strength may be insufficient to reflect spatially variant tissue properties and limit registration performance. In this study, we proposed to incorporate a spatially variant deformation prior into image registration framework using a statistical generative model.

Approach: A generator network is trained in an unsupervised setting to maximize the likelihood of observing the moving and fixed image pairs, using an alternating back-propagation approach. The trained model imposes constraints on deformation and serves as an effective low-dimensional deformation parametrization. During registration, optimization is performed over this learned parametrization, eliminating the need for explicit regularization and tuning. The proposed method was tested against SimpleElastix, DIRNet, and Voxelmorph.

Results: Experiments with synthetic images and simulated CTs showed that our method yielded registration errors significantly lower than SimpleElastix and DIRNet. Experiments with cardiac magnetic resonance images showed that the method encouraged physical and physiological feasibility of deformation. Evaluation with left ventricle contours showed that our method achieved a dice of (0.93  ±  0.03) with significant improvement over all SimpleElastix options, DIRNet, and VoxelMorph. Mean average surface distance was on millimeter level, comparable to the best SimpleElastix setting. The average 3D registration time was 12.78 s, faster than 24.70 s in SimpleElastix.

Conclusions: The learned implicit parametrization could be an efficacious alternative to regularized B-spline model, more flexible in admitting spatial heterogeneity.

Purpose: The segmentation of brain tumors is one of the most active areas of medical image analysis. While current methods perform superhuman on benchmark data sets, their applicability in daily clinical practice has not been evaluated. In this work, we investigate the generalization behavior of deep neural networks in this scenario.

Approach: We evaluate the performance of three state-of-the-art methods, a basic U-Net architecture, and a cascadic Mumford–Shah approach. We also propose two simple modifications (which do not change the topology) to improve generalization performance.

Results: In these experiments, we show that a well-trained U-network shows the best generalization behavior and is sufficient to solve this segmentation problem. We illustrate why extensions of this model in a realistic scenario can be not only pointless but even harmful.

Conclusions: We conclude from these experiments that the generalization performance of deep neural networks is severely limited in medical image analysis especially in the area of brain tumor segmentation. In our opinion, current topologies are optimized for the actual benchmark data set but are not directly applicable in daily clinical practice.

Purpose: The goal of this study was to quantify the effects of iterative reconstruction on radiomics features of normal anatomic structures on head and neck computed tomography (CT) scans.

Methods: Regions of interest (ROI) containing five different tissue types and an ROI containing only air were extracted from CT scans of the head and neck from 108 patients. Each scan was reconstructed using three different iDose⁴ reconstruction levels (2, 4, and 6) in addition to bone, thin slice (1-mm slice thickness), and thin-bone reconstructions. From each ROI in all reconstructions, 142 radiomic features were calculated. For each of the six ROIs, features were compared between combinations of iDose levels (2v4, 4v6, and 2v6) with a threshold of α  =  0.05 after correcting for multiple comparisons (p  <  0.00006). Features from iDose⁴-2 reconstructions were also compared to bone, thin slice, and thin-bone reconstructions. Spearman’s rank correlation coefficient, ρ, quantified the relative feature value agreement across iDose⁴ reconstructions.

Results: When comparing radiomics features across the three iDose⁴ reconstruction levels, over half of all features reflected significant differences for all tissue types, while no features demonstrated significant differences when extracted from air ROIs. When assessing feature value agreement, at least 97% of features reflected excellent agreement (ρ  >  0.9) when comparing the three iDose levels for all ROIs. When comparing iDose⁴-2 to bone, thin slice, and thin-bone reconstructions, more than half of all features demonstrated significant differences for all ROIs and 89  %   of features reflected excellent agreement for all ROIs.

Conclusion: Many radiomics features are dependent on the iterative reconstruction level, and the magnitude of this dependency is affected by the tissue from which features are extracted. For studies using images reconstructed using varying iDose⁴ reconstruction levels, features robust to these should be used.

Purpose: Prostate cancer (PCa) is the most common solid organ cancer and second leading cause of death in men. Multiparametric magnetic resonance imaging (mpMRI) enables detection of the most aggressive, clinically significant PCa (csPCa) tumors that require further treatment. A suspicious region of interest (ROI) detected on mpMRI is now assigned a Prostate Imaging-Reporting and Data System (PIRADS) score to standardize interpretation of mpMRI for PCa detection. However, there is significant inter-reader variability among radiologists in PIRADS score assignment and a minimal input semi-automated artificial intelligence (AI) system is proposed to harmonize PIRADS scores with mpMRI data.

Approach: The proposed deep learning model (the seed point model) uses a simulated single-click seed point as input to annotate the lesion on mpMRI. This approach is in contrast to typical medical AI-based approaches that require annotation of the complete lesion. The mpMRI data from 617 patients used in this study were prospectively collected at a major tertiary U.S. medical center. The model was trained and validated to classify whether an mpMRI image had a lesion with a PIRADS score greater than or equal to PIRADS 4.

Results: The model yielded an average receiver-operator characteristic (ROC) area under the curve (ROC-AUC) of 0.704 over a 10-fold cross-validation, which is significantly higher than the previously published benchmark.

Conclusions: The proposed model could aid in PIRADS scoring of mpMRI, providing second reads to promote quality as well as offering expertise in environments that lack a radiologist with training in prostate mpMRI interpretation. The model could help identify tumors with a higher PIRADS for better clinical management and treatment of PCa patients at an early stage.

Purpose: Hyperspectral imaging (HSI) is a non-contact optical imaging technique with the potential to serve as an intraoperative computer-aided diagnostic tool. Our work analyzes the optical properties of visible structures in the surgical field for automatic tissue categorization.

Approach: Building an HSI-based computer-aided tissue analysis system requires accurate ground truth and validation of optical soft tissue properties as these show large variability. We introduce and validate two different hyperspectral intraoperative imaging setups and their use for the analysis of optical tissue properties. First, we present an improved multispectral filter-wheel setup integrated into a fully digital microscope. Second, we present a novel setup of two hyperspectral snapshot cameras for intraoperative usage. Both setups are operating in the spectral range of 400 up to 975 nm. They are calibrated and validated using the same database and calibration set.

Results: For validation, a color chart with 18 well-defined color spectra in the visual range is analyzed. Thus the results acquired with both settings become transferable and comparable to each other as well as between different interventions. On patient data of two different otorhinolaryngology procedures, we analyze the optical behaviors of different soft tissues and show a visualization of such different spectral information.

Conclusion: The introduced calibration pipeline for different HSI setups allows comparison between all acquired spectral information. Clinical in vivo data underline the potential of HSI as an intraoperative diagnostic tool and the clinical usability of both introduced setups. Thereby, we demonstrate their feasibility for the in vivo analysis and categorization of different human soft tissues.

Purpose: In addition to less frequent and more comprehensive tests, quality assurance (QA) protocol for a magnetic resonance imaging (MRI) scanner may include cursory daily or weekly phantom checks to verify equipment constancy. With an automatic image analysis workflow, the daily QA images can be further used to study scanner baseline performance and both long- and short-term variations in image quality. With known baselines and variation profiles, automatic error detection can be employed.

Approach: Four image quality parameters were followed for 17 MRI scanners over six months: signal-to-noise ratio (SNR), image intensity uniformity, ghosting artifact, and geometrical distortions. Baselines and normal variations were determined. An automatic detection of abnormal QA images was compared with image deviations visually detected by human observers.

Results: There were significant inter-scanner differences in the QA parameters. In some cases, the results exceeded commonly accepted tolerances. Scanner field strengths, or a unit being stationary versus mobile, did not have a clear relationship with the QA results.

Conclusions: The variations and baseline levels of image QA parameters can differ significantly between MRI scanners. Scanner specific error thresholds based on parameter means and standard deviations are a viable option for detecting abnormal QA images.

Purpose: Alterations in the optical absorption behavior of liver tissue secondary to pathological processes can be evaluated by multispectral analysis, which is increasingly being explored as an imaging adjunct for use in liver surgery. Current methods are either invasive or have a limited wavelength spectrum, which restricts utility. This proof of concept study describes the development of a multispectral imaging (MSI) method called multispectral tissue mapping (MTM) that addresses these issues.

Approach: The imaging system consists of a tunable excitation light source and a near-infrared camera. Following the development stage, proof of concept experiments are carried out where absorption spectra from colorectal cancer liver metastasis (CRLM), hepatocellular carcinoma (HCC), and liver steatosis specimen are acquired and compared to controls. Absorption spectra are compared to histopathology examination as the current gold standard for tissue assessment. Generalized linear mixed modeling is employed to compare absorption characteristics of individual pixels and to select wavelengths for false color image processing with the aim of visually enhancing cancer tissue.

Results: Analysis of individual pixels revealed distinct absorption spectra therefore suggesting that MTM is possible. A prominent absorption peak at 1210 nm was found in lipid-rich animal tissues and steatotic liver specimen. Liver cancer tissue had a heterogeneous appearance on MSI. Subsequent statistical analysis suggests that measuring changes in absorption behavior may be a feasible method to estimate the pixel-based probability of cancer being present. In CRLM, this was observed throughout 1100 to 1700 nm, whereas in HCC it was concentrated around 1140 and 1430 nm. False color image processing visibly enhances contrast between cancer and normal liver tissues.

Conclusions: The system’s ability to enable no-touch MSI at 1100 to 1700 nm was demonstrated. Preliminary data suggest that MTM warrants further exploration as a potential imaging tool for the detection of liver cancer during surgery.

Purpose: In recent years, there has been increased clinical interest in the right ventricle (RV) of the heart. RV dysfunction is an important prognostic marker for several cardiac diseases. Accurate modeling of the RV shape is important for estimating the performance. We have created computationally effective models that allow for accurate estimation of the RV shape.

Approach: Previous approaches to cardiac shape modeling, including modeling the RV geometry, has used Doo–Sabin surfaces. Doo–Sabin surfaces allow effective computation and adapt to smooth, organic surfaces. However, they struggle with modeling sharp corners or ridges without many control nodes. We modified the Doo–Sabin surface to allow for sharpness using weighting of vertices and edges instead. This was done in two different ways. For validation, we compared the standard Doo–Sabin versus the sharp Doo–Sabin models in modeling the RV shape of 16 cardiac ultrasound images, against a ground truth manually drawn by a cardiologist. A Kalman filter fitted the models to the ultrasound images, and the difference between the volume of the model and the ground truth was measured.

Results: The two modified Doo–Sabin models both outperformed the standard Doo–Sabin model in modeling the RV. On average, the regular Doo–Sabin had an 8-ml error in volume, whereas the sharp models had 7- and 6-ml error, respectively.

Conclusions: Compared with the standard Doo–Sabin, the modified Doo–Sabin models can adapt to a larger variety of surfaces while still being compact models. They were more accurate on modeling the RV shape and could have uses elsewhere.

The erratum adds funding information missing from the Acknowledgments section of the originally published version of the article.