A domain knowledge-based interpretable deep learning system for improving clinical breast ultrasound diagnosis.

BACKGROUND: Though deep learning has consistently demonstrated advantages in the automatic interpretation of breast ultrasound images, its black-box nature hinders potential interactions with radiologists, posing obstacles for clinical deployment. METHODS: We proposed a domain knowledge-based interpretable deep learning system for improving breast cancer risk prediction via paired multimodal ultrasound images. The deep learning system was developed on 4320 multimodal breast ultrasound images of 1440 biopsy-confirmed lesions from 1348 prospectively enrolled patients across two hospitals between August 2019 and December 2022. The lesions were allocated to 70% training cohort, 10% validation cohort, and 20% test cohort based on case recruitment date. RESULTS: Here, we show that the interpretable deep learning system can predict breast cancer risk as accurately as experienced radiologists, with an area under the receiver operating characteristic curve of 0.902 (95% confidence interval = 0.882 - 0.921), sensitivity of 75.2%, and specificity of 91.8% on the test cohort. With the aid of the deep learning system, particularly its inherent explainable features, junior radiologists tend to achieve better clinical outcomes, while senior radiologists experience increased confidence levels. Multimodal ultrasound images augmented with domain knowledge-based reasoning cues enable an effective human-machine collaboration at a high level of prediction performance. CONCLUSIONS: Such a clinically applicable deep learning system may be incorporated into future breast cancer screening and support assisted or second-read workflows.

Breast cancer is one of the most common cancers, and finding it early can greatly improve patients' chances of survival and recovery. We create a tool based on artificial intelligence (AI)­whereby computer software learns to perform tasks that normally require human thinking­called MUP-Net. MUP-Net can analyze medical images to predict a patient's risk of having breast cancer. To make this AI tool usable in clinical practice, we enabled doctors to see the reasoning behind the AI's predictions by visualizing the key image features it analyzed. We showed that our AI tool not only makes doctors more confident in their diagnosis but also helps them make better decisions, especially for less experienced doctors. With further testing, our AI tool may help clinicians to diagnose breast cancer more accurately and quickly, potentially improving patient outcomes.

Noninvasive imaging-guided ultrasonic neurostimulation with arbitrary 2D patterns and its application for high-quality vision restoration.

Retinal degeneration, a leading cause of irreversible low vision and blindness globally, can be partially addressed by retina prostheses which stimulate remaining neurons in the retina. However, existing electrode-based treatments are invasive, posing substantial risks to patients and healthcare providers. Here, we introduce a completely noninvasive ultrasonic retina prosthesis, featuring a customized ultrasound two-dimensional array which allows for simultaneous imaging and stimulation. With synchronous three-dimensional imaging guidance and auto-alignment technology, ultrasonic retina prosthesis can generate programmed ultrasound waves to dynamically and precisely form arbitrary wave patterns on the retina. Neuron responses in the brain's visual center mirrored these patterns, evidencing successful artificial vision creation, which was further corroborated in behavior experiments. Quantitative analysis of the spatial-temporal resolution and field of view demonstrated advanced performance of ultrasonic retina prosthesis and elucidated the biophysical mechanism of retinal stimulation. As a noninvasive blindness prosthesis, ultrasonic retina prosthesis could lead to a more effective, widely acceptable treatment for blind patients. Its real-time imaging-guided stimulation strategy with a single ultrasound array, could also benefit ultrasound neurostimulation in other diseases.

Stretchable ultrasonic arrays for the three-dimensional mapping of the modulus of deep tissue.

Serial assessment of the biomechanical properties of tissues can be used to aid the early detection and management of pathophysiological conditions, to track the evolution of lesions and to evaluate the progress of rehabilitation. However, current methods are invasive, can be used only for short-term measurements, or have insufficient penetration depth or spatial resolution. Here we describe a stretchable ultrasonic array for performing serial non-invasive elastographic measurements of tissues up to 4 cm beneath the skin at a spatial resolution of 0.5 mm. The array conforms to human skin and acoustically couples with it, allowing for accurate elastographic imaging, which we validated via magnetic resonance elastography. We used the device to map three-dimensional distributions of the Young's modulus of tissues ex vivo, to detect microstructural damage in the muscles of volunteers before the onset of soreness and to monitor the dynamic recovery process of muscle injuries during physiotherapies. The technology may facilitate the diagnosis and treatment of diseases affecting tissue biomechanics.

Estimating Thrombus Elasticity by Shear Wave Elastography to Evaluate Ultrasound Thrombolysis for Thrombus With Different Stiffness.

OBJECTIVE: There is uncertainty about deep vein thrombosis standard treatment as thrombus stiffness alters each case. Here, we investigated thrombus' stiffness of different compositions and ages using shear wave elastography (SWE). We then studied the effectiveness of ultrasound-thrombolysis on different thrombus compositions. METHODS: Shear waves generated through mechanical shaker and traveled along thrombus of different hematocrit (HCT) levels, whereas 18-MHz ultrasound array used to detect wave propagation. Thrombus' stiffness was identified by the shear wave speed (SWS). In thrombolysis, a 3.2 MHz focused transducer was applied to different thrombus compositions using different powers. The thrombolysis rate was defined as the percentage of weight loss. RESULTS: The estimated average SWS of 20%, 40%, and 60% HCT thrombus were 0.75 m/s, 0.44 m/s, and 0.32 m/s, respectively. For Thrombolysis, the percentage weight loss at 8 MPa Negative pressure for the same HCT groups were 23.1%, 35.29%, and 39.66% respectively. CONCLUSION: SWS is inversely related to HCT level and positively related to thrombus age. High HCT thrombus had higher weight loss compared to low HCT. However, the difference between 20% and 40% HCT was more significant than between 40% and 60% HCT in both studies. Our results suggest that thrombus with higher SWS require more power to achieve the same thrombolysis rate as thrombus with lower SWS. SIGNIFICANCE: Characterizing thrombus elastic property undergoing thrombolysis enables evaluation of ultrasound efficacy for fractionating thrombus and reveals the appropriate ultrasound parameters selection to achieve a certain thrombolysis rate in the case of a specific thrombus stiffness.

Simultaneous Assessment of the Whole Eye Biomechanics Using Ultrasonic Elastography.

OBJECTIVE: Current elastography techniques in the field of ophthalmology usually target one specific tissue, such as the cornea or the sclera. However, the eye is an inter-related organ, and some ocular diseases can alter the biomechanical properties of multiple anatomical structures. Hence, there is a need to develop an imaging tool that can non-invasively, quantitatively, and accurately characterize dynamic changes among these biomechanical properties. METHODS: A high resolution ultrasound elastography system was developed to achieve this goal. The efficacy and accuracy of the system was first validated on tissue-mimicking phantoms while mechanical testing measurements served as the gold standard. Next, an in vivo elevated intraocular pressure (IOP) model was established in rabbits to further test our system. In particular, elastography measurements were obtained at 5 IOP levels, ranging from 10 mmHg to 30 mmHg in 5 mmHg increments. Spatial-temporal maps of the multiple ocular tissues (cornea, lens, iris, optic nerve head, and peripapillary sclera) were obtained. RESULTS: The spatial-temporal maps were acquired simultaneously for the ocular tissues at the 5 different IOP levels. The statistical analysis of the elastic wave speed was presented for ocular tissues. Finally, the mapping for the elastic wave speed of each ocular component was acquired at each IOP level. CONCLUSION: Our elastography system can concurrently assess the biomechanical properties of multiple ocular structures and detect changes in biomechanical properties associated with changes in IOP. SIGNIFICANCE: This system provides a novel tool to measure and quantify the biomechanical properties of the whole eye.

Ultrasound Retinal Stimulation: A Mini-Review of Recent Developments.

Ultrasound neuromodulation is an emerging technology. A significant amount of effort has been devoted to investigating the feasibility of noninvasive ultrasound retinal stimulation. Recent studies have shown that ultrasound can activate neurons in healthy and degenerated retinas. Specifically, high-frequency ultrasound can evoke localized neuron responses and generate patterns in visual circuits. In this review, we recapitulate pilot studies on ultrasound retinal stimulation, compare it with other neuromodulation technologies, and discuss its advantages and limitations. An overview of the opportunities and challenges to develop a noninvasive retinal prosthesis using high-frequency ultrasound is also provided.

High-Frequency Ultrasound Elastography to Assess the Nonlinear Elastic Properties of the Cornea and Ciliary Body.

Mechanical properties of the anterior anatomical structures of the eye, such as the cornea and ciliary body, play a key role in the ocular function and homeostasis. However, measuring the biomechanical properties of the anterior ocular structures, especially deeper structures, such as the ciliary body, remains a challenge due to the lack of high-resolution imaging tools. Herein, we implement a mechanical shaker-based high-frequency ultrasound elastography technique that can track the induced elastic wave propagation to assess the linear and nonlinear elastic properties of anterior ocular structures. The findings of this study advance our understanding of the role of anterior ocular structures in the pathogenesis of different ocular disorders, such as glaucoma.

Noninvasive Ultrasound Retinal Stimulation for Vision Restoration at High Spatiotemporal Resolution.

Objective. Retinal degeneration involving progressive deterioration and loss of function of photoreceptors is a major cause of permanent vision loss worldwide. Strategies to treat these incurable conditions incorporate retinal prostheses via electrically stimulating surviving retinal neurons with implanted devices in the eye, optogenetic therapy, and sonogenetic therapy. Existing challenges of these strategies include invasive manner, complex implantation surgeries, and risky gene therapy. Methods and Results. Here, we show that direct ultrasound stimulation on the retina can evoke neuron activities from the visual centers including the superior colliculus and the primary visual cortex (V1), in either normal-sighted or retinal degenerated blind rats in vivo. The neuron activities induced by the customized spherically focused 3.1 MHz ultrasound transducer have shown both good spatial resolution of 250 µm and temporal resolution of 5 Hz in the rat visual centers. An additional customized 4.4 MHz helical transducer was further implemented to generate a static stimulation pattern of letter forms. Conclusion. Our findings demonstrate that ultrasound stimulation of the retina in vivo is a safe and effective approach with high spatiotemporal resolution, indicating a promising future of ultrasound stimulation as a novel and noninvasive visual prosthesis for translational applications in blind patients.

Super-Resolution Ultrasound Localization Microscopy for Visualization of the Ocular Blood Flow.

OBJECTIVE: The ocular vascular system plays an important role in preserving the visual function. Alterations in either anatomy or hemodynamics of the eye may have adverse effects on vision. Thus, an imaging approach that can monitor alterations of ocular blood flow of the deep eye vasculature ranging from capillary-level vessels to large supporting vessels would be advantageous for detection of early stage retinal and optic nerve diseases. METHODS: We propose a super-resolution ultrasound localization microscopy (ULM) technique that can assess both the microvessel and flow velocity of the deep eye with high resolution. Ultrafast plane wave imaging was acquired using an L22-14v linear array on a high frequency Verasonics Vantage system. A robust microbubble localization and tracking technique was applied to reconstruct ULM images. The experiment was first performed on pre-designed flow phantoms in vitro and then tested on a New Zealand white rabbit eye in vivo calibrated to various intraocular pressures (IOP) - 10 mmHg, 30 mmHg and 50 mmHg. RESULTS: We demonstrated that retinal/choroidal vessels, central retinal artery, posterior ciliary artery, and vortex vein were all visible at high resolution. In addition, reduction of vascular density and flow velocity were observed with elevated IOPs. CONCLUSION: These results indicate that super-resolution ULM is able to image the deep ocular tissue while maintaining high resolution that is comparable with optical coherence tomography angiography. SIGNIFICANCE: Capability to detect subtle changes of blood flow may be clinically important in detecting and monitoring eye diseases such as glaucoma.

Prospective assessment of breast cancer risk from multimodal multiview ultrasound images via clinically applicable deep learning.

The clinical application of breast ultrasound for the assessment of cancer risk and of deep learning for the classification of breast-ultrasound images has been hindered by inter-grader variability and high false positive rates and by deep-learning models that do not follow Breast Imaging Reporting and Data System (BI-RADS) standards, lack explainability features and have not been tested prospectively. Here, we show that an explainable deep-learning system trained on 10,815 multimodal breast-ultrasound images of 721 biopsy-confirmed lesions from 634 patients across two hospitals and prospectively tested on 912 additional images of 152 lesions from 141 patients predicts BI-RADS scores for breast cancer as accurately as experienced radiologists, with areas under the receiver operating curve of 0.922 (95% confidence interval (CI) = 0.868-0.959) for bimodal images and 0.955 (95% CI = 0.909-0.982) for multimodal images. Multimodal multiview breast-ultrasound images augmented with heatmaps for malignancy risk predicted via deep learning may facilitate the adoption of ultrasound imaging in screening mammography workflows.

High resolution optical coherence elastography of retina under prosthetic electrode.

BACKGROUND: Quantitatively investigating the biomechanics of retina with a retinal prosthetic electrode, we explored the effects of the prosthetic electrode on the retina, and further supplemented data for a potential clinical trial. METHODS: Biomechanical properties were assessed with a high resolution optical coherence tomography (OCT) based elastography (OCE) system. A shaker was used to initiate elastic waves and an OCT system was used to track axial displacement along with wave propagation. Rabbits received surgery to implant the retinal prosthetic electrode, and elastic wave speed was measured before and after implantation; anatomical B-mode images were also acquired. RESULTS: Spatial-temporal maps of each layer in retina with and without prosthetic electrodes were acquired. Elastic wave speed of nerve fiber to inner plexiform layer, inner nuclear to outer nuclear layer, retinal pigmented epithelium layer and choroid to sclera layer without prosthetic electrode were found to be 3.66±0.36, 5.33±0.07, 6.85±0.37, and 9.69±0.24 m/s, respectively. With prosthetic electrode, the elastic wave speed was found to be 4.09±0.26, 5.14±0.11, 6.88±0.70, and 9.99±0.73 m/s, respectively in each layer. CONCLUSIONS: Our results show that the elastic wave speed in each layer of retina is slightly faster with the retinal electrode, and further demonstrate that the retinal prosthetic electrode does not affect biomechanical properties significantly. In the future, we expect OCE technology to be used by clinicians where it could become part of routine testing and evaluation of the biomechanical properties of the retina in response to long term use of prosthetic electrodes in patients.

Ultrasonic elastography to assess biomechanical properties of the optic nerve head and peripapillary sclera of the eye.

PURPOSE: To quantitatively investigate both optic nerve head (ONH) and peripapillary sclera (PPS) biomechanical properties of porcine eyes through an ultrasonic elastography imaging system in response to both increasing and decreasing intraocular pressure (IOP). METHODS: The Young's modulus of the ONH and PPS were assessed using our high resolution ultrasonic imaging system which utilized a mechanical shaker to induce shear waves and an off-axis aligned 40 MHz needle transducer to track micron-level displacement along the direction of wave propagation. In this study, imaging on a total of 8 ex vivo porcine eyes preloaded with IOPs from 6 mmHg to 30 mmHg was performed. To have a better understanding of the effect of varying IOP on biomechanics, both increasing and decreasing IOPs were investigated. RESULTS: The increase of the Young's modulus of ONH (92.4 ± 13.9 kPa at 6 mmHg to 224.7 ± 71.1 kPa at 30 mmHg) and PPS (176.8 ± 14.3 kPa at 6 mmHg to 573.5 ± 64.4 kPa at 30 mmHg) following IOP elevation could be observed in the reconstructed Young's modulus of the shear wave elasticity (SWE) imaging while the B-mode structural images remained almost unchanged. In addition, for the same IOP level, both ONH and PPS have a tendency to be stiffer with decreasing IOP as compared to increasing IOP. CONCLUSIONS: Our results demonstrate the feasibility of using our ultrasonic elastography system to investigate the stiffness mapping of posterior eye with high resolution in both increasing and decreasing IOPs, making this technique potentially useful for glaucoma.

Layer-specific ultrasound elastography using a multi-layered shear wave dispersion model for assessing the viscoelastic properties.

Mechanical properties of biological tissues are significant biomarkers for diagnosing various diseases. Assessing the viscoelastic properties of multi-layer tissues has remained challenging for a long time. Some shear wave models have been proposed to estimate thin-layer tissues' viscoelasticity recently. However, the potential applications of these models are highly restricted since few biological tissues are single-layered. Here we proposed a multi-layer model for layer-specific viscoelasticity estimation of biological tissues. Integrating the theoretical model and ultrasonic micro-elastography imaging system, the viscoelasticity of both layers was assessed. Dual-layer phantoms and ex vivo porcine eyes were used to verify the proposed model. Results obtained from the mechanical test and shear wave rheological model using bulk phantoms were provided as validation criteria. The representative phantom had two layers with elastic moduli of 1.6 ± 0.2 kPa and 18.3 ± 1.1 kPa, and viscosity moduli of 0.56 ± 0.16 Pa·s and 2.11 ± 0.28 Pa·s, respectively. The estimated moduli using the proposed model were 1.3 ± 0.2 kPa and 16.20 ± 1.8 kPa, and 0.80 ± 0.31 Pa·s and 1.87 ± 0.67 Pa·s, more consistent with the criteria (one-tailed t-test, p < 0.1). By contrast, other methods, including the group velocity method and single-layer Rayleigh-Lamb model, generate significant errors in their estimates. For the ex vivo porcine eye, the estimated viscoelasticity was 23.2 ± 8.3 kPa and 1.0 ± 0.4 Pa·s in the retina, and 158.0 ± 17.6 kPa and 1.2 ± 0.4 Pa·s in the sclera. This study demonstrated the potential of the proposed method to significantly improve accuracy and expand clinical applications of shear wave elastography.

2-D Ultrasonic Array-Based Optical Coherence Elastography.

Acoustic radiation force optical coherence elastography (ARF-OCE) has been successfully implemented to characterize the biomechanical properties of soft tissues, such as the cornea and the retina, with high resolution using single-element ultrasonic transducers for ARF excitation. Most currently proposed OCE techniques, such as air puff and ARF, have less capability to control the spatiotemporal information of the induced region of deformation, resulting in limited accuracy and low temporal resolution of the shear wave elasticity imaging. In this study, we propose a new method called 2-D ultrasonic array-based OCE imaging, which combines the advantages of 3-D dynamic electronic steering of the 2-D ultrasonic array and high-resolution optical coherence tomography (OCT). The 3-D steering capability of the 2-D array was first validated using a hydrophone. Then, the combined 2-D ultrasonic array OCE system was calibrated using a homogenous phantom, followed by an experiment on ex vivo rabbit corneal tissue. The results demonstrate that our newly developed 2-D ultrasonic array-based OCE system has the capability to map tissue biomechanical properties accurately, and therefore, has the potential to be a vital diagnostic tool in ophthalmology.

Transcranial Focused Ultrasound for Noninvasive Neuromodulation of the Visual Cortex.

Currently, blindness cannot be cured and patients' living quality can be compromised severely. Ultrasonic (US) neuromodulation is a promising technology for the development of noninvasive cortical visual prosthesis. We investigated the feasibility of transcranial focused ultrasound (tFUS) for noninvasive stimulation of the visual cortex (VC) to develop improved visual prosthesis. tFUS was used to successfully evoke neural activities in the VC of both normal and retinal degenerate (RD) blind rats. Our results showed that blind rats showed more robust responses to ultrasound stimulation when compared with normal rats. ( , two-sample t-test). Three different types of ultrasound waveforms were used in the three experimental groups. Different types of cortical activities were observed when different US waveforms were used. In all rats, when stimulated with continuous ultrasound waves, only short-duration responses were observed at "US on and off" time points. In comparison, pulsed waves (PWs) evoked longer low-frequency responses. Testing different parameters of PWs showed that a pulse repetition frequency higher than 100 Hz is required to obtain the low-frequency responses. Based on the observed cortical activities, we inferred that acoustic radiation force (ARF) is the predominant physical mechanism of ultrasound neuromodulation.

From Light to Sound: Photoacoustic and Ultrasound Imaging in Fundamental Research of Alzheimer's Disease.

Alzheimer's disease (AD) causes severe cognitive dysfunction and has long been studied for the underlining physiological and pathological mechanisms. Several biomedical imaging modalities have been applied, including MRI, PET, and high-resolution optical microscopy, for research purposes. However, there is still a strong need for imaging tools that can provide high spatiotemporal resolutions with relatively deep penetration to enhance our understanding of AD pathology and monitor treatment progress in fundamental research. Photoacoustic (PA) imaging and ultrasound (US) imaging can potentially address these unmet needs in AD research. PA imaging provides functional information with endogenous and/or exogenous contrast, while US imaging provides structural information. Recent studies have demonstrated the ability to monitor physiological parameters in small-animal brains with PA and US imaging as well as the feasibility of using US imaging as a therapeutic tool for AD. This concise review aims to introduce recent advances in AD research using PA and US imaging, provide the fundamentals, and discuss the potentials and challenges for future advances.

Current Ultrasound Technologies and Instrumentation in the Assessment and Monitoring of COVID-19 Positive Patients.

Since the emergence of the COVID-19 pandemic in December of 2019, clinicians and scientists all over the world have faced overwhelming new challenges that not only threaten their own communities and countries but also the world at large. These challenges have been enormous and debilitating, as the infrastructure of many countries, including developing ones, had little or no resources to deal with the crisis. Even in developed countries, such as Italy, health systems have been so inundated by cases that health care facilities became oversaturated and could not accommodate the unexpected influx of patients to be tested. Initially, resources were focused on testing to identify those who were infected. When it became clear that the virus mainly attacks the lungs by causing parenchymal changes in the form of multifocal pneumonia of different levels of severity, imaging became paramount in the assessment of disease severity, progression, and even response to treatment. As a result, there was a need to establish protocols for imaging of the lungs in these patients. In North America, the focus was on chest X-ray and computed tomography (CT) as these are widely available and accessible at most health facilities. However, in Europe and China, this was not the case, and a cost-effective and relatively fast imaging modality was needed to scan a large number of sick patients promptly. Hence, ultrasound (US) found its way into the hands of Chinese and European physicians and has since become an important imaging modality in those locations. US is a highly versatile, portable, and inexpensive imaging modality that has application across a broad spectrum of conditions and, in this way, is ideally suited to assess the lungs of COVID-19 patients in the intensive care unit (ICU). This bedside test can be done with little to no movement of the patients from the unit that keeps them in their isolated rooms, thereby limiting further exposure to other health personnel. This article presents a basic introduction to COVID-19 and the use of the US for lung imaging. It further provides a high-level overview of the existing US technologies that are driving development in current and potential future US imaging systems for lung, with a specific emphasis on portable and 3-D systems.

In Vivo Visualization of Eye Vasculature Using Super-Resolution Ultrasound Microvessel Imaging.

OBJECTIVE: The choroidal vessels, which supply oxygen and nutrient to the retina, may play a pivotal role in eye disease pathogenesis such as diabetic retinopathy and glaucoma. In addition, the retrobulbar circulation that feeds the choroid shows an important pathophysiologic role in myopia and degenerative myopia. Owing to the light-absorbing retinal pigment epithelium (RPE) and optically opaque sclera, choroidal and retrobulbar vasculature were difficult to be observed using clinically accepted optical coherence tomography angiography (OCT-A) technique. Here, we have developed super-resolution ultrasound microvessel imaging technique to visualize the deep ocular vasculature. METHODS: An 18-MHz linear array transducer with compounding plane wave imaging technique and contrast agent - microbubble was implemented in this study. The centroid intensity of each microbubble was detected using image deconvolution algorithm with spatially variant point spread function, and then accumulated in successive frames in order to reconstruct microvasculature. The image deconvolution technique was first evaluated in a simulation study and experimental flow phantoms. The performance was then validated on normal rabbit eyes in vivo. RESULTS: The image deconvolution based super-resolution ultrasound microvessel imaging technique shows good performance on either simulation study or flow phantoms. In vivo rabbit eye study indicated that the micron-level choroidal and retrobulbar vessels around the optic nerve head were successfully reconstructed in multiple 2D views and 3D volume imaging. CONCLUSIONS: Our results demonstrate the capability of using super-resolution ultrasound microvessel imaging technique to image the microvasculature of the posterior pole of the eye. This efficient approach can potentially lead to a routinely performed diagnostic procedure in the field of ophthalmology.

A combined ultrasonic B-mode and color Doppler system for the classification of breast masses using neural network.

OBJECTIVES: To develop a dual-modal neural network model to characterize ultrasound (US) images of breast masses. MATERIALS AND METHODS: A combined US B-mode and color Doppler neural network model was developed to classify US images of the breast. Three datasets with breast masses were originally detected and interpreted by 20 experienced radiologists according to Breast Imaging-Reporting and Data System (BI-RADS) lexicon ((1) training set, 103212 masses from 45,433 + 12,519 patients. (2) held-out validation set, 2748 masses from 1197 + 395 patients. (3) test set, 605 masses from 337 + 78 patients). The neural network was first trained on training set. Then, the trained model was tested on a held-out validation set to evaluate agreement on BI-RADS category between the model and the radiologists. In addition, the model and a reader study of 10 radiologists were applied to the test set with biopsy-proven results. To evaluate the performance of the model in benign or malignant classifications, the receiver operating characteristic curve, sensitivities, and specificities were compared. RESULTS: The trained dual-modal model showed favorable agreement with the assessment performed by the radiologists (κ = 0.73; 95% confidence interval, 0.71-0.75) in classifying breast masses into four BI-RADS categories in the validation set. For the binary categorization of benign or malignant breast masses in the test set, the dual-modal model achieved the area under the ROC curve (AUC) of 0.982, while the readers scored an AUC of 0.948 in terms of the ROC convex hull. CONCLUSION: The dual-modal model can be used to assess breast masses at a level comparable to that of an experienced radiologist. KEY POINTS: • A neural network model based on ultrasonic imaging can classify breast masses into different Breast Imaging-Reporting and Data System categories according to the probability of malignancy. • A combined ultrasonic B-mode and color Doppler neural network model achieved a high level of agreement with the readings of an experienced radiologist and has the potential to automate the routine characterization of breast masses.