This PDF file contains the front matter associated with SPIE Proceedings Volume 11602 including the Title Page, Copyright information, and Table of Contents.

Welcome and Introduction to SPIE Medical Imaging conference 11602: Ultrasonic Imaging and Tomography

Automatic delineation of relevant structures in intravascular imaging can support percutaneous coronary interventions (PCIs), especially when dealing with rather demanding cases. We found three major error types which occur regularly when segmenting lumen and wall of morphologically complex vessels with convolutional neural networks (CNNs). In order to reduce these three error types, we developed three IVUS-specific methods which are able to improve generalizability of state-of-the-art CNNs for IVUS segmentation tasks. These methods are based on three concepts: speckle statistics, artery shape priors via independent component analysis (ICA) and the concentricity condition of lumen and vessel wall. We found that all three methods outperform the baseline. Since all three concepts can be readily transferred to intravascular optical coherence tomography (IVOCT), we expect these findings can support the segmentation of corresponding images as well.

The aim of this study is to identify ultrasound texture features to characterize radiation-associated acute breast toxicity in women following radiotherapy treatment for breast cancer. We investigated a series of sonographic features obtained from the gray level co-occurrence matrix (GLCM) – a second order statistical method of texture analysis. These features were tested in a pilot study of 42 postradiotherapy patients. The mean follow-up time for the postradiotherapy patients was 6 weeks. Each participant underwent an ultrasound study in which ultrasound scans were performed on the bilateral breast, generating a total of 42 post irradiation and 42 contralateral normal breasts exams. The ultrasound scans of the irradiated breasts were graded as either mild (n=27) or severe (n=15). After specifying the region of interest on the B-mode images and computing the sonographic features based off the severity grading, we observed statistically significant differences in the quantification of

Recently, Convolutional Neural Networks (CNNs) have been very successful in optical flow estimation in computer vision. UltraSound Elastography (USE) displacement estimation step can be performed by optical flow CNNs. However, there is a large domain gap between ultrasound Radio-Frequency (RF) data and computer vision images which reduces the overall accuracy of displacement estimation. Some modifications of the network architecture are required to be able to extract reliable information from RF data. Modified Pyramidal Network (MPWC-Net) which is based on the well-known PWC-Net was among the first attempts that adopts the optical flow CNNs to USE displacement estimation. However, MPWC-Net suffers from several shortcomings that limit its application especially for unsupervised training. In this paper, we propose additional modifications to substantially improve MPWC-Net. We also publicly released the network’s trained weights.

In this paper, a limited-angle Ultrasound Computed Tomography (USCT) system capturing quantitative features is presented. Quantitative characteristics of tissues such as speed of sound (SS) and attenuation coefficient (AC), have great potential to distinguish malignant and benign tissues. The proposed system requires two facing linear array transducers to measure the time of flight and the amplitude of traversed waves. The Quantitative Imaging Network (QI-Net) is modeled and trained for stable image reconstruction from ultrasonic information achieved from two facing limited-angle probes. In addition, a Quantitative Imaging Network incorporating a priori information (QIP-Net) to the neural network is also presented. A robust ROI compression scheme embedded in the proposed networks extracts quantitative image information regardless of the measurement size. We evaluated our methods via numerical simulation, phantom, and ex-vivo measurements. The simulation results show that the QI-Net and QIP-Net are capable of quantifying SS with the average error of 1.1m/s (0.56%) and 4.5m/s (2.3%), respectively. In the phantom and ex-vivo studies, the networks demonstrate accurate extraction of SS and AC under diverse conditions.

In full waveform inversion (FWI) for ultrasound computed tomography (CT), choosing the right sound source is essential for generating high-resolution images. We developed an optimized source estimation method for FWI to efficiently reduce the value of any cost function and evaluated its performance using simulation data and measurement data. In our optimized source estimation method, we obtain the sound source as α(k)·f, where f is a sound source (an arbitrary complex value), coefficient α(k) is cos(θ_k)+i·sin(θ_k), k is integer (0, 1, ⋯, N-1), θ_k is 2π/N·k, and the integer value of N is 360. We then determine the coefficient α(k) that minimizes the value of the cost function. In contrast to conventional source estimation, which only minimizes the value of the L2 norm cost function, our proposed source estimation can minimize the values of any cost function, such as the L1 norm or a hybrid of L1 and L2 norms. The advantage of our method is that it can be easily applied to FWI with various cost functions. In this preliminary study, we implemented FWI with the L2 norm cost function and compared the performance of our proposed method with that of the conventional method. In the simulation study, FWI with both the conventional and proposed source estimation methods improved the contrasts of inclusions of a numerical phantom compared to FWI with no source estimation. They both also improved the contrasts of inclusions of a measured oil-gel-based phantom compared to a bent-ray reconstruction method. The absolute mean errors between ROI and true values were 39, 11, and 11 [m/s] for the bent-ray reconstruction method, FWI with the conventional method, and FWI with the proposed method, respectively. In addition, FWI with both the conventional and proposed methods improved the contrasts of a patient’s tumor compared to the bentray reconstruction method. These results demonstrate that FWI with the proposed source estimation method can provide the same contrast and quantitative accuracy as FWI with the conventional source estimation method.

Ultrasound transmission tomography promises a high potential and novel imaging method for early breast cancer diagnosis; it can quantitatively characterize tissues or materials by the attenuation and speed of sound (SoS). Reconstruction of ultrasound transmission tomography is an inverse problem that can be solved iteratively based on a paraxial approximation of the Helmholtz equation as forward model, which is highly non-linear and time-consuming. In order to address these problems and reconstruct desired images, we design a dual domain network architecture for ultrasound transmission tomography reconstruction. It can enhance the information of measurement domain and directly reconstruct from pressure field measurements without using any initialization of reconstruction and fully connected layer. We train the network on simulated ImageNet data and transfer it for ultrasound transmission tomography images to avoid overfitting when the amount of ultrasound transmission tomography images is limited. Our experimental results demonstrate that a dual domain network produces significant improvements over state-of-the-art methods. It improves the measured structural similarity measure (SSIM) from 0.54 to 0.90 and normalized root mean squared error (nRMSE) from 0.49 to 0.01 on average concerning the SoS reconstruction, and from 0.46 to 0.98 for SSIM, from 353 to 0.03 for nRMSE on average concerning the attenuation reconstruction.

The backscatter coefficient (BSC) quantifies the frequency-dependent reflectivity of tissues. Accurate estimation of the BSC requires knowledge of the attenuation coefficient slope (ACS) of tissues in the beam path between the transducer and the insonified region of interest, namely, the total attenuation. In this study, the total attenuation is calculated as the cumulative sum of values of a local attenuation map devised using full angular spatial compounding (FASC). The BSC was parameterized through the integrated backscatter coefficient (iBSC) obtaining iBSC maps. Experimental validation of the proposed approach consisted of scanning two cylindrical physical phantoms with off-centered inclusions having different ACS and BSC values than the background. Additional iBSC maps were computed assuming an uniform ACS map of 0.5 dB/cm/MHz (which is a value assumed for soft tissues) instead of the FASC-ACS map. Finally a iBSC map was obtained using an ideal ACS map formed with ground truth ACS values and knowledge of inclusion true position. The results were comparable when using the FASC-ACS map or the ideal ACS map in term of inclusion detectability and estimation accuracy. The use of the uniform ACS map resulted in some cases with very high fractional error (>;9 dB), which highlights the relevance of accurate compensation for total attenuation. These results suggest that BSCs can be reliably estimated using total attenuation compensation from FASC-ACS maps.

We propose here a real time approach for aberration correction and Speed Of Sound quantification in the context of ultrafast ultrasound imaging. Theoretical works demonstrate the physical meaning of Singular Value Decomposition of the Ultrafast Compound Matrix. In heterogeneous media, we show that the first spatial and angular singular vectors retrieve respectively the non-aberrated image, and the aberration law. In vitro and in vivo experiments probe the efficiency of our method for angular coherence optimization through severe aberrations. Besides correcting the image, it provides knowledge of phase and amplitude aberration, allowing also a local Speed Of Sound estimation.

The advancement in bio-engineering technology has enabled tissues to be artificially cultivated from human cells, providing the opportunity to model disease and discover potential treatments 1 . Blood vessel is an important category of human tissues that can be artificially engineered to facilitate the development of treatment plans for vascular diseases. The growth of tissue engineered blood vessels (TEBVs) is a costly procedure, and effective quality control during the growing process could help reduce waste and optimize the cultivation process. Imaging technologies, such as optical coherence tomography5,6 (OCT), have been applied to obtain cross-sectional images of TEBVs, which could be used as a nondestructive method to assess blood vessel during cultivation. Ultrasound (US) imaging has been widely accepted in clinical practice due to its real-time imaging capacity and zero radiation emission; and compared to optics-based imaging modality it is more accessible financially. We implemented an US computer tomography (USCT) based monitoring system on assisting quality control in TEBV growth. In this prototype, a single element transducer is placed in a circular stand that rotates around the TEBV bioreactor to collect A-lines from different angles. Mechatronics systems are used to actuate the transducer for circular motion. A circular back-projection method is used in image reconstruction. Experiments were carried out with point phantom and the bioreactor to validate the imaging functionality of the prototype. Reconstructed images provide validation to the feasibility of using USCT to monitor the growth of TEBV growth.

Osteoarthritis (OA) is the most common chronic health condition and a leading cause of disability and pain in the United States and Canada. Current methods for monitoring the development of knee OA (KOA) involve x-ray radiography and magnetic resonance imaging (MRI) to assess femoral articular cartilage (FAC) degradation and synovial membrane inflammation. However, x-ray radiography cannot be used to quantify FAC loss or synovial membrane inflammation due to a lack of soft tissue contrast, and MRI is associated with high costs, long waitlists, long scan times, and is inaccessible to many patients. We have developed a counterbalanced point-of-care (POC) system to track multiple three-dimensional (3D) ultrasound (US) acquisitions and register them for visualizing the entire suprapatellar synovium. This work aims to validate the tracking accuracy of the POC system against an external optical tracking system. Validation was conducted using optical tracking as a reference by mounting a custom optical tracking stylus in place of the 3D US scanner on the POC system. The stylus was manipulated in 3D space and Euclidean distances were calculated using the initial and final positions of the stylus and were compared between systems to quantify POC tracking system error. Results indicated that the overall mean absolute tracking error of the POC system was 3.08 ± 2.01 mm with no statistically significant difference between the POC and optical systems (p = 0.965). The POC system has the potential to enable clinicians and researchers to obtain additional information without added complexity or discomfort to patients.

Mammographic quantitative breast density (QBD), the ratio of fibroglandular tissue to whole breast volume, is known to be important for risk assessment for breast cancer. Most methods are based on 2D projections, though some use MRI. We show two methods for determining QBD from 3D ultrasound tomographic (UT) images, their equivalence and superiority over other methods of estimation. False assignments to breast density can occur if projection methods are used. A sigmoidal function is fit using a log likelihood maximization and the QBD from MRI images is compared with QBD as calculated from 3D UT showing strong correlation.

Limited options exist to improve oxygenation in patients with acute hypoxic respiratory failure. We are developing an intravascular oxygenator catheter capable of delivering a clinically significant amount of O2 directly into the bloodstream. A critical design consideration is minimizing the risk of embolism. We demonstrate the use of ultrafast ultrasound imaging, coherence beamforming, and image processing to quantify bubble formation in a benchtop flow circuit. We found no significant increase in bubble counts throughout 30 minutes of oxygenation using the device.

3D inverse scattering ultrasound tomography (3D UT) is quantitative and not subject to artifacts from 2D algorithms and data and does not require contrast agents or ionizing radiation. However, it is time consuming, so it is important to have timing results for 3D inverse scattering reconstructions of the whole breast with 3D algorithms and full 3D data and in the clinically relevant context of a diverse population of dense, heterogeneously dense and fatty breasts. The adaptive algorithm uses different reconstruction frequencies and iteration counts for different breasts. We compare a computational complexity count with the observed fit of reconstruction times vs breast size that and show performance comparable to published TFLOP performance for nVidia cards. We show a reconstruction time of 24 minutes for an average size breast and show substantial speed up with more efficient nVidia cards. These numbers indicate clinical viability for 3D transmission ultrasound even in the clinical setting with diverse demographics in low income areas. The cohort of 23 cases of different types of breasts were reconstructed on two P6000's and compared with the same data reconstructed on two RTX6000's with 24GB on-board memory and some optimization of the CUDA code. The resulting speed up is better than linear with increasing computation time, indicating increasing efficiency with computational complexity, larger breasts. Image quality is also affected, since increasing iteration and frequency counts give generally better images as long as overconvergence is avoided. These results further validate the 3D Quantitative UT as clinically viable, especially for underserved populations.

High frequency ultrasound biomicroscopy (UBM) images are used in clinical ophthalmology due to its ability to penetrate opaque tissues and create high resolution images of deeper intraocular structures. Because these inexpensive, high frequency (50 MHz) systems use single ultrasound elements, there is a limitation in visualizing small structures and anatomical landmarks, especially outside focal area, due to the lack of dynamic focusing. The wide and axially variant point spread function degrade image quality and obscure smaller structures. We created a fast, generative adversarial network (GAN) method to apply axially varying deconvolution for our 3D ultrasound biomicroscopy (3D-UBM) imaging system. Original images are enhanced using a computationally expensive axially varying deconvolution, giving paired original and enhanced images for GAN training. Supervised generative adversarial networks (pix2pix) were trained to generate enhanced images from originals. We obtained good performance metrics (SSIM = 0.85 and PSNR = 31.32 dB) in test images without any noticeable artifacts. GAN deconvolution runs at about 31 msec per frame on a standard graphics card, indicating that near real time enhancement is possible. With GAN enhancement, important ocular structures are made more visible.

In this work, a system is presented for combined therapy planning, low intensity FUS treatment, and real-time therapy monitoring using a single diagnostic imaging array. First, a sonication pattern was determined by manually segmenting the treatment region from a B-mode image captured with the imaging array. To visualize the FUS therapy beam, a focused pulse excitation was transmitted and backscattered signals were used to reconstruct the intensity field of the FUS beam. The FUS beam reconstruction was overlaid onto a co-aligned B-mode image captured with the imaging array, allowing one to qualitatively monitor the position and size of the FUS beam with anatomical context from the B-mode image. Real-time beam visualizations at a frame rate of 25-30 frames per second were achieved in a rat tumor in vivo and a mock FUS therapy was planned and monitored in a tissue-mimicking.

Brachytherapy (BT) is a form of radiation therapy which typically relies on the insertion of needles to deliver radiation and is commonly used to treat a variety of cancer sites including prostate cancer. Accurate needle tip identification is key for safe and effective BT, as errors can result in radiation delivery that deviates from the planned dose. Typically, standard brightness (B)-mode ultrasound (US) imaging is used, however, artifacts can limit the visibility of needles. We propose a novel wireless mechanical oscillator for needle perturbation in a power Doppler (PD)-based needle tip identification method to overcome these limitations and improve needle tip identification accuracy in BT. We evaluated our method using a tissue equivalent phantom with the clinical US system and needles. In this proof-of-concept study, we assessed the performance of our PD US method using tip error computed based on a reference needle. Signed mean ± SD tip error was -0.03 ± 0.49 mm, while absolute mean error was 0.33 ± 0.33 mm, demonstrating our PD US method provided accurate needle tip identification for ideal visualization in phantom. We also demonstrated that our PD US method is capable of visualizing needles that cannot be seen using B-mode US due to shadowing, providing promise for the clinical utility of our method. Our wireless mechanical oscillator for PD-based needle tip identification is simple to use and easy to implement, requiring no modifications to clinical equipment. A prospective clinical trial has been approved to assess our method in the clinic.

MRI and Transrectal Ultrasound (TRUS) prostate images provide essential information for prostate intervention and cancer treatment. It is beneficial to register the MRI with TRUS images for cancer diagnosis and tumor delineation. However, the two imaging modalities have distinct image intensity which makes the registration difficult. A deep learning based image registration framework was proposed to perform MRI-TRUS image fusion. The prostate from the MRI and TRUS was first segmented for prostate shape modeling using point clouds. The network utilizes point cloud matching to perform prostate alignment. Biomechanical modeling was used to regularize the prostate deformation during image registration. The performance of the network was evaluated using dice similarity coefficient (DSC), mean surface distance (MSD), Hausdorff distance (HD) and target registration error (TRE). The calculated DSC, MSD, HD and TRE were 0.94±0.02, 0.90±0.23mm, 2.96±1.00mm and 1.57±0.77mm, respectively. Our results showed that the proposed method can register the prostate on MRI to that on the TRUS accurately

Photoacoustic imaging (PAI) can be used to infer molecular information about myocardial health non-invasively in vivo using optical excitation at ultrasonic resolution. For clinical and preclinical linear array imaging systems, conventional delay-and-sum (DAS) beamforming is typically used. However, DAS is prone to image quality degradation when applied to murine cardiac PAI resulting in low signal specificity in the myocardium. To address this, we propose a spatiotemporal singular value decomposition (SVD) processing method using electrocardiogram (ECG) and respiratory gated in vivo cardiac murine PAI data. SVD was applied on a two-dimensional spatiotemporal matrix generated using a threedimensional volume of DAS beamformed complex PAI data over a cardiac cycle. The singular value spectrum (SVS) was then filtered to remove contributions from static clutter and random noise. Finally, SVD processing of beamformed images were derived using filtered SVS and inverse SVD computations. In vivo murine cardiac PAI was performed by collecting single wavelength (850 nm) photoacoustic channel data using two healthy mice. Qualitative comparison with DAS shows that SVD processed images had better signal specificity and contrast. DAS and SVD processed PAI were quantitatively evaluated by calculating contrast ratio (CR), generalized contrast-to-noise ratio (gCNR) and signal-to-noise ratio (SNR). SVD processed PAI had higher CR, gCNR and SNR values compared to DAS results. For example, at the end-systolic phase for mouse 1, the SVD processed image had 100.48% higher gCNR compared to the DAS image. These results suggest that significantly better-quality images can be realized using spatiotemporal SVD processing for in vivo murine cardiac PAI.

In this report, we explore the strengths and limitations of principal component analysis (PCA) and independent component analysis (ICA) for clutter and noise filtering in ultrasonic peripheral perfusion imaging. The advantages of pre-filtering spatial registration to reduce the bandwidth of coherent clutter motion is also considered. PCA methods excel when the echo covariance exhibits a significant blood-scattering component orthogonal to the tissue clutter component. This situation exists in peripheral perfusion imaging when the echo signals are temporally stationary and normally distributed. ICA methods separate non-orthogonal blood-clutter echo components often found in moving clutter, but only for echo signals with either non-normal-amplitude distributions or nonstationary normal distributions. When clutter movement is large and spatially coherent, echo registration followed by PCA filtering can be ideal. Effective filtering is essential for contrast-free ultrasonic perfusion imaging of muscle tissues in the extremities of patients at risk for developing peripheral artery diseases. Statistical filter performance is examined using simulation and echo data from an in vivo ischemic hindlimb mouse model.

Immediate postoperative assessment of trans-arterial chemoembolization (TACE) using gold standard modalities, MRI and CT, is unreliable due to confounding interactions with lipiodol and post-embolization inflammatory changes. We previously demonstrated that recent advancements in power Doppler ultrasound processing enables changes in slow blood flow to be detected immediately following TACE. Recently, we have developed a filtering method that employs a higherorder singular value decomposition (HOSVD) applied to aperture data to mitigate thermal noise and acoustic clutter. Here, we investigate HOSVD as a tool to improve non-contrast ultrasound evaluation of TACE. Preliminary feasibility is demonstrated in a small pilot study. Treatment-induced changes in perfusion are visualized most readily using the HOSVD filter in comparison to conventional filtering methods. The HOSVD filter produced the greatest change in contrast between pre-TACE and post-TACE power Doppler images.

Non-contrast ultrasound blood flow imaging is difficult at slow blood flow rates. Singular value decomposition (SVD) and independent component analysis (ICA) are useful for separating tissue, blood, and noise sources for Doppler filtering. In addition, it has been shown that applying SVD and ICA in a block-wise manner further improves source separation; noise within a small block is theoretically more stationary, and thus easier to separate. Yet, there is much discussion on how to select independent components; several methods have been introduced with some success. We present a novel, adaptive hierarchical clustering approach for selecting appropriate independent components for blood flow image filtering that utilizes Kurtosis and Normalized Cross Correlation. Components are clustered based on the Kurtosis and NCC of each component; an optimal number of clusters is chosen using the Silhouette Method. Appropriate clusters are selected based on the Autocorrelation Function of each cluster. Our method was tested on 1 mm/s and 5 mm/s flowrate phantoms containing a 0.6 mm vessel and resulted in average SNR and CNR increases of 6.2 dB and 3.7 dB, respectively, for 1 mm/s blood flow velocities. We demonstrate that our method improves tissue and noise suppression throughout the field of view while maintaining blood flow information.

Ultrasound power Doppler imaging is a useful clinical tool for measuring perfusion. Sensitivity to slow moving blood flow is important for many clinical applications, but thick abdominal walls or the presence of bone such as ribs or the skull cause significant attenuation and thereby reduce the signal-to-noise ratio (SNR) and flow sensitivity. One way to improve SNR is to inject microbubble contrast agents into the vascular system, but this is impractical for many applications. An alternative approach is to use coded excitation, a signal processing technique that can drastically increase SNR within FDA safety limits without contrast agents. This work encompasses a method to design long coded pulses that are simple to implement along with a pulse compression technique to completely suppress range lobes, thereby recovering axial resolution, maintaining contrast, and improving SNR by as much as a factor of 10log10(code length). In simulations we show that this approach reliably improves the SNR of power Doppler imaging across a range of noise levels. As the noise level increases with respect to the blood, contrast and contrast-to-noise ratio are maintained with coded excitation whereas they drop precipitously without coded excitation. In vivo feasibility is also shown in transcranial and transthoracic cardiac B-Mode imaging. Both simulation and in vivo results match theoretical expectations of SNR gain. Finally, preliminary results showing in vivo power Doppler imaging in the liver are presented as well. Coded excitation is able to improve the blood vessel to background CNR and CR as compared to a standard approach.

Reverberation clutter is a difficult source of image degradation in patients, and minimum variance (MV) in particular is poorly equipped to handle such sources, as we demonstrate here. We propose that a pre-processing step such as ADMIRE should be implemented in cases with high reverberation clutter when we still want to be able to implement MV to realize improvements in lateral resolution. The ADMIRE, or aperture domain model image reconstruction, method is specifically designed to suppress or eliminate reverberant and off-axis sources of clutter while returning the decluttered channel data with its original dimensionality, allowing us to sequentially process the data with MV. We show that in simulated data this combined method results in clear improvements to image quality, contrast ratio, and target boundary thickness compared to DAS and MV alone. In in vivo cases, contrast ratio and general image quality are improved, and boundary thickness is generally on par with DAS and MV.

Ultrasound Strain imaging (USI) is a radiofrequency (RF) signal-based method for mapping mechanical tissue properties with widespread preclinical and clinical uses. USI quality is contingent upon the accuracy of estimated displacement fields and incorporation of regularization has significantly improved it. Here, we report on a Bayesian spatiotemporal regularization (ST-Bayes) scheme which estimates displacement using four consecutive RF frames. ST-Bayes iteratively regularizes 2-D normalized cross-correlation (NCC) metrics incorporating information from adjacent spatial and temporal neighbors in a Bayesian framework. Regularized NCC metrics are posterior probability density-derived using likelihood and NCC as prior and integrated into a three-level block matching (BM) method. Algorithm is validated using inclusion phantom data acquired under free-hand compression and mouse common carotid artery (MCCA) datasets collected using high-frequency transducers. ST-Bayes was compared against NCC and spatial-regularization (S-Bayes) method. For the inclusion phantom, ST-Bayes provided strain images with improved lesion boundary and background noise reduction for both RF and RF + noise data (SNR_s = 10 dB) compared to NCC and S-Bayes. ST-Bayes improved CNR_e by 7.32 % and 62.08 % when compared to S-Bayes and NCC, respectively, for RF, and by 17.17 % and 219.43 % for RF + noise data. ST-Bayes also provided smoother displacement curves in MCCA, reducing strain variance, indicating robust regularization using spatiotemporal information.

Ultrasound (US) is a low-cost, portable, and safe tool for breast cancer screening. However, automatic classification of invasive ductal carcinoma (IDC) in US is a difficult classification task due to their similar appearance to fibroadenoma (FA) (a type of benign tumor). Another challenge is the limited availability of US data with ground truth labels, further complicating the adoption of deep learning techniques for IDC detection. It has been shown that deep classification networks perform better when they simultaneously learn multiple correlated tasks. However, most previous studies on breast US classifications focused on the binary classification of benign versus malignant tumors. To this end, we propose a multi-class classification deep learning-based strategy mainly focusing on the classification of IDC. Inspired by multi-task learning (MTL), we adopt a novel scheme in adding the background tissue as an additional class and show substantial improvements in IDC detection.

Ultrasound computed tomography offers to be a transformative breast imaging technology; however, quantitative measurement and artefact-free reconstruction of the diagnostically important property of acoustic attenuation has proven to be a challenge. The UK’s National Physical Laboratory (Teddington, UK) have developed a prototype ultrasound computed tomography system which utilises a novel sensor technology specifically designed for quantitative acoustic attenuation tomography. This scanning system, with its phase-insensitive receiver, has been used to generate 2D and 3D quantitative images of the acoustic attenuation of commercially-available and bespoke breast phantoms made by CIRS (Norfolk, VA, USA). The acoustic attenuation measurements and images obtained from the system were evaluated in comparison to acoustically characterised phantom materials and Xray computed tomography imagery. The images, while generated using a relatively fast and simple reconstruction method, were found to be less susceptible to artefacts common to acoustic attenuation tomography.

The intensity of backscattered ultrasound signal from heart tissue is known to be related to the angle between cardiac fibers and the insonification direction. In this work, we use optical ray tracing, fast convolutions with the system point spread function, and an empirically derived relationship between backscatter intensity and fiber orientation to simulate plane wave echocardiography on the GPU. First, we simulate grayscale images of a rotating fiber phantom, and validate that the angle-to-backscatter relationship is accurately reflected by the simulated radiofrequency data. Second, we use our method to simulate view-dependent echocardiography images of human heart tissue from diffusion tensor magnetic resonance imaging (DT-MRI) data. Simulated backscatter intensity measurements show excellent agreement with the underlying angle-to-backscatter relationship, and echocardiography images exhibit view-dependent speckle with a realistic appearance. Our GPU-based simulation method is fast and generates biologically realistic images, making it particularly useful for fiber orientation quantification and many other ultrasound imaging studies.

In silico studies for ultrasound computed tomography (USCT) can allow to explore imaging system parameters and reconstruction methods, without the economic burden and ethical concerns of clinical trials. A framework is proposed for virtual imaging trials of USCT. First, an ensemble of three-dimensional numerical breast phantoms consisting of anatomically realistic tissue structures and lesions is created. Next, acoustic properties are assigned to each tissue-type within physiological ranges. Finally, USCT measurement data are computed by simulating acoustic wave propagation. The proposed framework will establish a standard pipeline for USCT virtual imaging trials and provide publicly available large-scale datasets

This work presents a method to estimate average anisotropy in speed of sound of soft tissue using pulse-echo ultrasound. In particular, our setup includes a passive acoustic reflector located opposite to the ultrasound probe, with tissue in between. This enables the generation of strong reflections from which we measure their traveltimes. We use ray-based approaches to derive the forward problem that relates observed traveltimes with speed of sound anisotropy parameters. The accuracy of the forward modelling is verified using numerical wave propagation simulations. We finally show the occurrence of anisotropy in muscle tissue using in-vivo data.

Recently, X-ray-induced acoustic computed tomography (XACT) has been raised as a promising imaging modality that can monitor the dose delivery in radiation therapy by combining high x-ray absorption contrast with the 3D propagation advantages provided by high-resolution ultrasound waves. k-Wave toolbox is a powerful MATLAB toolbox that can simulate the generation, propagation, and detection of the thermal acoustic wave. The purpose of this study was to summarize the k-Wave simulation works on XACT imaging. In our group, two simulation experiments were carried out, focusing on determining the radiation dose in stereotactic partial breast irradiation (SPBI) therapy and prostate radiation therapy, respectively. The simulation results demonstrated that the k-Wave simulation tool has great potential in simulating wave propagation and reconstructing XACT dose images according to initial pressure distribution.

Ultrasound computed tomography (USCT) is a promising imaging modality for breast cancer screening. Two challenges commonly arising in time-of-flight USCT are (1) to efficiently deal with large data sets and (2) to effectively mitigate the ill-posedness for an adequate reconstruction of the model. In this contribution, we develop an optimization strategy based on a stochastic descent method that adaptively subsamples the data, and analyze its performance in combination with different sparsity-enforcing regularization techniques. The algorithms are tested on numerical as well as real data obtained from synthetic phantom scans of the previous USCT Data Challenges.

Ultrasound computer tomography (USCT) is a promising modality for breast cancer diagnosis which images the reflectivity, sound speed and attenuation of tissue. Elastic properties of breast tissue, however, cannot directly be imaged although they have shown to be applicable as a discriminator between different tissue types. In this work we propose a novel approach combining USCT with the principles of strain elastography. Socalled USCT-SE makes use of imaging the breast in two deformation states, estimating the deformation field based on reconstructed images and thereby allows localizing and distinguishing soft and hard masses. We use a biomechanical model of the breast to realistically simulate both deformation states of the breast. The analysis of the strain is performed by estimating the deformation field from the deformed to the undeformed image by a non-rigid registration. In two experiments the non-rigid registration is applied to ground truth sound speed images and simulated SAFT images. Results of the strain analysis show that for both cases soft and hard lesions can be distinguished visually in the elastograms. This paper provides a first approach to obtain mechanical information based on external mechanical excitation of breast tissue in a USCT system.

Overdiagnosis and overtreatment are two major risks involved with mammography-based breast screening which, in addition to its 3D variant, is currently the only approved breast screening technology. Quantitative Transmission (QT) ultrasound is an upcoming breast imaging modality that has the ability to generate quantitative speed-of-sound based maps of the whole breast enabling unprecedented imaging biomarkers. On top of that, machine learning (ML)-based methods for breast tissue/cancer classification have shown promise because of their unique advantages in innovative feature mining from complex datasets. In this paper, we present the results of the deployment of novel data-driven, yet explainable methods implemented in Albeado’s PRISM AI/ML platform that delivers rapid and accurate breast mass detection when applied to QT imaging. Using PRISM, we first deployed computer vision methods to segment breast tissue and identify regions of interest (ROI) for the three-dimensional volumetric speed-of-sound maps, which allows for further classification into benign and malignant masses using unsupervised methods. Our strategy is to segment breast images into candidate units, extract radiomic features for each unit, and then distinguish normal tissue from pathological tissue. In order to evaluate our lesion detection framework, we rank the lesions according to their radiomic features and compare the top ranking candidates to radiologist annotations. For malignant cases, lesions are consistently identified (95% recall). Our results indicate that the presented radiomics-based method is a viable candidate for breast mass detection and classification in QT imaging and serves as a framework for further development.

Full-waveform inversion applied to ultrasound computed tomography is a promising technique to provide highresolution quantitative images of soft human tissues, which are otherwise difficult to illuminate by conventional ultrasound imaging. A particular challenge which arises within transcranial ultrasound is the imprint of the solid skull on the measured wavefield. We present an acoustoelastic approach to full-waveform inversion for transcranial ultrasound computed tomography that accurately accounts for the solid-fluid interactions along the skull-tissue interfaces. Using the spectral-element method on cubical meshes, we obtain a scalable and performant method to resolve such a coupled physical system. Moreover, since the volume of the skull is small compared to the entire simulation domain, solving a coupled system of the acoustoelastic wave equation increases the computational cost only by a small margin compared to the acoustic approximation. We perform an in silico forward and inverse modeling study that reveals significant coupling effects at the skull-tissue interfaces when considering the skull as an elastic medium as opposed to an acoustic medium. Applying full-waveform inversion to a set of synthetically generated acoustoelastic forward data allows for favorable reconstructions to be achieved when considering an acoustoelastic prior model of the skull.

The ultrasound computed tomography system based on the ring probe adopts sequential single-slice scanning mode. The ring probe keeps stationary when the scan is in process; after the signals of one slice are captured, the ring probe steps along the elevation axis to scan the next slice. The scanning process needs to step and stop repeatedly, and the scanning time is relatively long. Also, the lack of signals between tomographic image slices may result in missed diagnosis. This paper proposes a new spiral synthetic aperture method for ultrasound tomographic volume imaging. The ring probe moves along the elevation axis at a certain speed, while the transmitting and receiving array elements are switched electronically. Therefore, “spiral” transmitting aperture and receiving aperture are formed, and continuous three-dimensional spatial spiral data are collected to directly reconstruct the three-dimensional volume image. Preliminary experimental result shows that this method can shorten the scanning time and improve image quality in the elevation direction.

Freehand 3D-ultrasound imaging using a 2D-ultrasound probe with attached position and orientation(pose) sensor is a cost-effective 3D-imaging modality. The standard method of 3D-reconstruction involves stacking the 2D-image slices in their appropriate position and orientations in the region swept by the probe. A single 2D-image is obtained by sequentially exciting the multiple elements of an array transducer probe. Standard 3D-reconstruction assumes that all the scanlines of a single frame are acquired concurrently. This assumption limits the speed of the scan required with accurate volume reconstruction. For correcting imperfections in reconstructed volume due to fast probe-motion, this paper proposes a new scanline-based 3D-volume reconstruction. This method corrects the placement and orientation of each scanline, such that the 'fast' probe motion does not distort the reconstructed volume. The improved performance of the proposed reconstruction method compared to standard plane-based reconstruction is demonstrated with scans performed with a convex probe on imaging phantoms. An anechoic cylindrical inclusion in a cube-shaped phantom is visualized using both reconstruction methods and visually compared to a reference image. Reference image of inclusion is reconstructed from a slow scan. For quantitative analysis, edge detected inclusions from images of the reconstructed volumes are compared with that of the corresponding reference image. The Dice coefficient of the inclusion with the proposed scanline-based reconstruction with the reference image is on an average 83% higher than that of the standard method.

Fractional flow reserve (FFR) is the reference standard to identify flow-limiting coronary stenosis that requires revascularization. Accurate computation of FFR from coronary intravascular images is based on the precise reconstruction of the side branches. In this paper, a novel approach for segmentation of side branches in intravascular images is presented. The framework consists of an image-to-image translation module and two side branch region segmentation modules. By using the image-to-image translation module, information from intravascular optical coherence tomography (IVOCT) and intravascular ultrasound (IVUS) images is combined to improve the segmentation performance. The framework is trained on a total of 62475 IVOCT and 186110 IVUS images, and evaluated on an independent dataset which contains 9344 IVOCT images with 91 side branches and 39450 IVUS images with 128 side branches. The Dice coefficients of IVOCT and IVUS side branches segmentation are 0.935±0.039 and 0.856±0.038, respectively. The validation results of side branches detection are: Precision = 0.934, Recall = 0.923, F₁Score = 0.929 in IVOCT, and 0.925, 0.868, 0.895 in IVUS, accordingly. Ablation studies demonstrate excellent efficiency in incorporating multi-modal information with our proposed image-to-image translation module.

Ultrasound imaging during or prior to radiation therapy offers a great potential in terms of safety, cost and real time imaging capacity. However, this task is challenging for tumors of abdominal such as liver cancer due to respiratory motion. In this work, we proposed an unsupervised deep-learning-based method to track the respiratory motion for 3D ultrasound (US) liver imaging. A Markov-like network, which extract features from consistent 3D US frames, was utilized to estimate a sequence of deformation vector fields (DVFs) that register tracked frame with landmark to match the coming untracked frames without landmark. Then, the landmarks of the coming frames were tracked by moving landmark position of tracked frame to the coming frames based on DVFs. The Markov-like network aims to consider motion consistency between each two frames and is implemented via a generative adversarial network (GAN). The proposed method was evaluated on the MICCAI CLUST 2015 challenge dataset. A retrospective study was performed with a total of 8 sets of 3D US sequence, and each 3D sequence has 4-96 frames marked with landmarks. We used a leave-one-out experiment to test our tracking method. Quantitative evaluation was performed by calculating the tracking error between estimated landmarks and the ground truth landmarks on each frame. Results of the proposed method showed a mean tracking error of 2.01±1.06 mm for 3D liver US images. We proposed a deep learning-based approach for 3D US liver motion tracking. This method reduces the processing time for tracking respiratory motion significantly, which can reduce the delivery uncertainty.

Digitalizing all the needles in ultrasound (US) images is a crucial step of treatment planning for US-guided high-dose-rate (HDR) prostate brachytherapy. However, current computer-aided technologies are broadly focused on single-needle digitization, while manual digitization of all needles is labor intensive and time consuming. In this paper, we proposed a deep learning-based workflow for fast automatic multi-needle digitization, including needle shaft detection and needle tip detection. The major workflow is composed of two components: a large margin mask R-CNN model (LMMask R-CNN), which adopts the lager margin loss to reformulate Mask R-CNN for needle shaft localization, and a needle-based density-based spatial clustering of application with noise (DBSCAN) algorithm which integrates priors to model a needle in an iteration for a needle shaft refinement and tip detections. Besides, we use the skipping connection in neural network architecture to improve the supervision in hidden layers. Our workflow was evaluated on 23 patients who underwent USguided HDR prostrate brachytherapy with 339 needles being tested in total. Our method detected 98% of the needles with 0.0911±0.0427 mm shaft error and 0.3303±0.3625 mm tip error. Compared with only using mask R-CNN and only using LMMask R-CNN, the proposed method gains a significant improvement of accuracy on both shaft and tip localization. The proposed method automatically digitizes needles per patient with in a second. It streamlines the workflow of USguided HDR prostate brachytherapy and paves the way for the development of real-time treatment planning system that is expected to further elevate the quality and outcome of HDR prostate brachytherapy.

For 3D ultrasound (US) images with large slice thickness, high frequency information in the slice direction is missing and cannot be resolved through interpolation. As an ill-posed problem, current high-resolution methods rely on the presence of external/training atlases to learn the transform from low resolution images to high resolution images. In this study, we aim to propose a self-supervised learning method, which does not use any external atlas images, yet can still resolve high resolution images only reliant on the acquired image with a large slice thickness. To circumvent the lack of training data, the simulated training data were obtained from the input image. To do this, each 2D sagittal slice is regarded as a high-resolution image, while each coronal and axial slice is regarded as low-resolution images. By training a deep learning-based model on sagittal slices and using this model to infer high-resolution coronal and axial slices, we can apply the mapping to low-resolution images with large slice thickness to estimate the high-resolution images with thin slice thickness. The proposed algorithm was evaluated using 30 sets of US breast data. The US image downsampled in z-axis was used as low-resolution image, the original US image was used as ground truth. The normalized mean absolute error (NMAE), peak signal-to-noise ratio (PSNR) and normalized cross-correlation (NCC) indices were used to quantify the accuracy of the proposed algorithm. The NMAE, PSNR and NCC were 0.011±0.02, 34.6±2.14 dB and 0.98±0.01. The proposed method showed similar image quality as compared to ground truth.

Automatic breast ultrasound (ABUS) imaging provides complementary information when other imaging modalities (i.e., mammography) are not conclusive in the task of tumor detection. It enables sectional plane visualization, simplified temporal comparison and coronal depiction, and features higher reproducibility than conventional ultrasound imaging. Although the 3D ABUS acquisition significantly reduces the acquisition time and cost, the manual segmentation of tumor on 3D ABUS could be time-consuming and labor-intensive due to its high slice number. This work aims to develop a deep-learning-based method to automatically segment the breast tumor on 3D ABUS. We integrated mask scoring-based self-correlation strategy into the R-CNN-based method to force the final segmented tumor contour to be more reasonable. We tested the performance of the proposed method using 3D ABUS of 40 patients who are confirmed with breast tumor through four-fold cross validation test. The comparison between our results and the ground truth contours was quantified by metrics including the Dice similarity coefficient (DSC), Jaccard index (JAC), 95% Hausdorff distance (HD95). The mean DSC, JAC, and HD95 were 0.85 ± 0.10, 0.75 ± 0.14, and 1.65 ± 1.20 (mm), respectively.

Intravascular ultrasound (IVUS) image is widely used in coronary atherosclerotic plaque analysis. The delineation of coronary lumen borders and external elastic lamina (EEL) in IVUS images is a crucial step in the analysis. Conventional segmentation approaches by delineating each image frame independently would lead to incorrect segmentation when guide wires, side branches, and calcified plaques presented. In this work, we proposed a new framework via mask propagation method for the segmentation of IVUS image sequences by using information from previous frame and current frame to predict the mask probability map. Experiments showed that our proposed method achieved an encouraging result with Dice Similarity Coefficient (Dice) of 0.973 and 0.949, Jaccard Index (JI) of 0.949 and 0.906, and Hausdorff Distance (HD) of 0.187 mm and 0.225 mm for delineation of EEL and lumen, respectively.

Cerebrovascular and cardiovascular diseases such as stroke and coronary artery disease show a significant number of cases with a high mortality rate. Early detection of risk factors is important to prevent cerebrovascular and cardiovascular diseases. Measurements of carotid artery stenosis, blood flow rate, and the wall thickness of vessels by 2D and Doppler mode ultrasound are preferred choices due to advantages of their easy access, non-invasiveness, and safety. However, the current ultrasound imaging system with handheld type probes is not suitable for continuous monitoring and imaging, and the manual measurement is required by qualified personnel at a given time interval. Therefore, it is not an ideal solution for collecting continuous time-series data. We developed a 32-element, patch-type linear array transducer with a small footprint of 11.73 mm x 8 mm, which is an acceptable size to be attached over the carotid artery in the neck area. We evaluated the performance of the developed array transducer using the pulse-echo system and obtained its representative center frequency of 4.5 MHz, bandwidth (-6 dB) 64%, and sensitivity -47 dB. We also implemented a compact tabletop ultrasound system capable of 2D-mode real-time imaging of carotid artery and Doppler measurement of blood flow. In addition, with the tissue-mimicking phantom, we evaluated the performance of the developed system by collecting 2D images and Doppler spectrogram. The -6 dB lateral resolutions of the ultrasound system were 0.76, 0.61, and 1.33 mm at 15, 25, and 35 mm, respectively, and the peak velocity of the Doppler signal was close to 100 cm/s.

Cotton balls are a versatile and efficient tool commonly used in neurosurgical procedures to absorb fluids and manipulate delicate tissues. However, the use of cotton balls is accompanied by the risk of accidental retention in the brain after surgery. Retained cotton balls can lead to dangerous immune responses and potential complications, such as adhesions and textilomas. In a previous study, we showed that ultrasound can be safely used to detect cotton balls in the operating area due to the distinct acoustic properties of cotton compared with the acoustic properties of surrounding tissue. In this study, we enhance the experimental setup using a 3D-printed custom depth box and a Butterfly IQ handheld ultrasound probe. Cotton balls were placed in variety of positions to evaluate size and depth detectability limits. Recorded images were then analyzed using a novel algorithm that implements recently released YOLOv4, a state-of-the-art, real-time object recognition system. As per the radiologists’ opinion, the algorithm was able to detect the cotton ball correctly 61% of the time, at approximately 32 FPS. The algorithm could accurately detect cotton balls up to 5mm in diameter, which corresponds to the size of surgical balls used by neurosurgeons, making the algorithm a promising candidate for regular intraoperative use.

The needle intervention is required in many clinical procedures such as lumbar puncture and lymph node biopsy. Ultrasound (US) imaging has been applied widely to guide procedures involving needle insertion. However, conventional 2D US image guidance provides a limited field of view (FOV) for the region of interest (ROI), especially toward the outof-plane axis. Also, a high level of hand-eye coordination is required to accurately place the needle to the target in tissue because the needle path and the image plane are not registered to each other. This paper p roposes a needle insertion mechanism that enables 3D US imaging to provide a larger FOV and thus assists clinicians in making better decisions about where to insert the needle. Besides 3D imaging functionality, the image plane and needle path are co-registered in the proposed mechanism to facilitate needle insertion. The proposed mechanism uses actuated acoustic reflectors to redirect acoustic waves to different parts of ROI. 2D image slices are collected along the elevation direction and are then focused on the elevation plane. Elevation focusing is achieved with a synthetic aperture focusing algorithm that considers the acoustic path geometry in the actuated reflector system. Both software simulation and imaging experiments using a prototype are carried out to validate the 3D imaging performance.The simulation and experiment results of point and wire phantoms validate the 3D imaging capa bility and suggest that image quality on the elevation plane after applying elevation focusing improved in terms of both resolution and signal to noise ratio. By providing clinicians with an extended FOV with improved image quality, the proposed mechanism has the potential to enable needle insertion with better efficiency and a higher success rate.

Calcium signaling is a second messenger that triggers physiological changes at the cellular level, such as proliferation, differentiation, or apoptosis. It was recently found that calcium dynamics play a vital role in many studies, including cancer, Alzheimer’s disease, and Parkinson’s disease, and we studied how intracellular signaling pathways work by ultrasound mechanotransduction. However, since ultrasound mechanotransduction does not yet have many experimental results by the quantified ultrasound parameters, little is known about the mechanism between ultrasound parameters and calcium dynamics. We investigate calcium level changes using different frequencies of ultrasound to study intracellular signal pathways of fibroblasts, which may function as one of the contributing factors of tissue repair. We quantified a few major ultrasound stimulation parameters, i.e., operating frequency, beam width, and acoustic pressure. Three 40 MHz ultrasound transducers with different f-numbers (0.8, 1.0, and 1.5) were designed and fabricated. During the cell stimulation, ultrasound waves with different frequencies (36, 45, and 69 MHz) but the same beam width and same acoustic pressure were exerted on the cells. The cell lines used were NIH/3T3 fibroblasts. At the fixed acoustic pressure and beam width, intracellular calcium level increased more rapidly at higher frequencies, which shows that the intracellular signal pathways of fibroblasts may be mainly dependent upon the frequency used for stimulation.

Photoacoustic imaging is a new non-destructive medical imaging technology based on photoacoustic effect. It can reflect the difference of light absorption energy by detecting photoacoustic signal. At present, the analysis methods of photoacoustic signals in biological tissues can be divided into two categories, namely, time-domain analysis of signals and frequency-domain analysis of signals. In time domain analysis, the envelope of the received photoacoustic signal is usually used to reconstruct the image. However, due to the influence of various external factors, the time domain signal cannot accurately reflect the characteristics of the absorber itself. Here, photoacoustic spectrum analysis was performed by using k-Wave to obtains the relationship between the structure, size, density of the absorber and the photoacoustic spectrum. Firstly, the relationship between the size of absorber and the photoacoustic spectrum is studied, and the slope and intercept are used to analyze the spectrum. Conversely, the relationship was used to predict the size of the absorber Finally, we used this relationship to predict the size of blood vessels.