Breast cancer is the leading cause of cancer worldwide and the second leading cause of cancer death among women. Ultrasound (US) is one of the most widely used tools for breast cancer screening and diagnosis - particularly for women with dense breast tissue where mammography performs less effectively. It is non-invasive, does not use radiation, is portable, and can guide biopsies in real time.
Despite these advantages, ultrasound has a critical limitation: it is highly operator-dependent. The quality and accuracy of results depend heavily on the skill and experience of the individual performing the scan. This leads to high false-negative rates - missed cancers - and inconsistent performance across different healthcare settings and geographic regions.
In underserved regions with limited access to specialist radiologists, delayed diagnosis of breast masses is a common and serious problem. The need for tools that can reduce dependence on expert interpretation while maintaining high accuracy has driven significant research into AI-assisted ultrasound analysis.
This review surveys research published between 2017 and early 2023 on applying deep learning models to breast ultrasound across the full spectrum of clinical use - from initial screening and lesion classification to predicting lymph node metastasis and monitoring responses to chemotherapy. It analyzes which ultrasound modes and which deep learning architectures have been used, and where they perform best.
Modern breast ultrasound encompasses several distinct imaging modes. B-mode (conventional grayscale imaging) is by far the most commonly used and was the basis for over 40 of the 59 studies reviewed. It produces real-time images of breast tissue structure and is the standard tool for characterizing masses found on mammography.
Shear Wave Elastography (SWE) measures tissue stiffness - a characteristic that can help distinguish benign from malignant masses, as cancerous tissue tends to be stiffer. Color Doppler and Power Doppler imaging visualize blood flow within and around lesions. Malignant tumors often develop abnormal vascularity, making Doppler a useful supplementary tool.
Automated Breast Ultrasound (ABUS) reduces operator variability by standardizing image acquisition. 3D ultrasound adds depth information to conventional 2D images. These multi-modal approaches have the potential to improve sensitivity significantly compared to B-mode alone, but their integration with AI has been underexplored.
The review found that the vast majority of deep learning studies (over 40 of 59) used only B-mode images, with multimodal combinations studied in far fewer cases. This means significant potential value from combining imaging modes with deep learning remains largely untapped.
The most commonly used deep learning architectures in breast ultrasound research are established convolutional neural networks (CNNs). AlexNet, VGG16/VGG19, ResNet, and GoogLeNet/Inception were the most frequently applied models for classification tasks. These architectures excel at extracting image features hierarchically - from simple edges to complex patterns that characterize malignant tissue.
For segmentation tasks - precisely outlining tumor boundaries within ultrasound images - U-Net was the dominant architecture. U-Net uses an encoder-decoder structure: the encoder compresses the image into learned feature representations, and the decoder reconstructs a pixel-level map identifying which regions belong to the tumor. Attention U-Net extends this by adding attention layers that help the network focus on diagnostically important areas like tumor margins.
Faster R-CNN was used for detection and classification tasks, combining a region proposal network with a CNN backbone to simultaneously locate tumors and classify them. DenseNet (which densely connects layers so each receives features from all preceding layers) and Xception were used for more complex tasks like predicting lymph node status and tumor molecular subtype.
Generative Adversarial Networks (GANs) were applied in a smaller number of studies, primarily for data augmentation - generating synthetic training images to address the challenge of limited labeled medical datasets. Across all architectures, the Adam optimizer was the most common training choice, followed by stochastic gradient descent, with cross-entropy as the predominant loss function.
The most studied application of deep learning in breast ultrasound is lesion classification - determining whether a detected breast mass is benign or malignant. This task showed the highest performance across all reviewed studies, with some models approaching 100% accuracy. The majority of these studies compared model performance to that of radiologists, with many deep learning systems matching or exceeding specialist performance.
Five studies focused on BI-RADS classification - categorizing lesions on a standard scale from 1 (normal) to 5 (highly suspicious) that guides clinical decision-making. One study using Xception achieved accuracy, sensitivity, and specificity of 89.7%, 91.3%, and 86.9% respectively for BI-RADS categorization. BI-RADS category 3 lesions (probably benign) showed the highest accuracy at near 100%, while category 4B (intermediate suspicion) was harder to classify correctly.
Segmentation - precisely delineating the boundaries of breast masses in ultrasound images - was the second most studied application. U-Net-based models achieved Dice coefficients (a measure of overlap between predicted and actual boundaries) of 80 to 97%, depending on the dataset and architecture variant. Accurate segmentation is clinically important for measuring tumor size, tracking treatment response, and planning surgery.
One notable challenge across classification and segmentation tasks is distinguishing between benign subtypes - such as fibroadenoma, adenosis, and papillary tumors. Malignant types like invasive ductal carcinoma showed classification accuracy of around 90%, but rarer subtypes including invasive lobular carcinoma and certain inflammatory lesions were more difficult, sometimes showing false-negative results on ultrasound regardless of AI model used.
Axillary lymph node (ALN) metastasis - the spread of cancer cells from the primary breast tumor to nearby lymph nodes - is one of the most important prognostic factors in breast cancer. Traditional assessment requires surgical sampling, which is invasive. Seven of the 59 reviewed studies applied deep learning to predict lymph node status directly from breast ultrasound images, providing a potential non-invasive alternative.
The most effective architectures for this task were DenseNet, ResNet, Mask R-CNN, and Xception. One study achieved 95% sensitivity, 96% specificity, and 98% accuracy in predicting ALN status from B-mode images. Another using DenseNet-121 achieved an AUC of 0.81. Performance was generally lower than for classification tasks, reflecting the greater complexity of predicting metastatic spread from primary tumor images alone.
Three studies investigated using deep learning to predict response to neoadjuvant chemotherapy - drugs given before surgery to shrink a tumor. Being able to predict or monitor treatment response from ultrasound would allow oncologists to adjust therapy earlier, sparing patients from ineffective treatment regimens. ResNet and VGG19 were used in these studies, with promising but preliminary results.
The review highlighted a significant gap: far fewer studies have been conducted on prognosis and treatment monitoring than on initial diagnosis and classification. This means the potential of deep learning to support the entire continuum of breast cancer management - not just its earliest stages - remains substantially unexplored.
The most commonly cited limitation across all 59 reviewed studies is small dataset size. Medical imaging datasets are inherently limited by data availability, privacy constraints, and the cost of expert annotation. Small datasets increase the risk of overfitting - where models perform well on training data but fail to generalize to new patients or settings.
Single-center data is another pervasive limitation. Models trained on images from one hospital's ultrasound equipment may not transfer well to different machines, transducer frequencies, or scanning protocols. This is compounded by the fact that most studies were retrospective - analyzing historical records rather than prospectively collecting data for the specific purpose of training AI models.
The black box problem affects clinical trust in deep learning models. Because decisions emerge from millions of learned parameters, the reasoning behind a specific classification cannot easily be explained to a radiologist or patient. Two FDA-authorized automated breast ultrasound systems (Smart Ultrasound by Koios and QVCAD by QViewMedical) have been developed, but broader clinical adoption requires better interpretability tools.
Methodological inconsistency across studies also makes comparison difficult. Performance metrics used vary widely (accuracy, AUC, Dice coefficient, sensitivity, specificity), hyperparameters are often poorly documented, and many studies do not report computational time - a key factor for real-world clinical deployment. The lack of standardized benchmarks makes it difficult to identify which architectures truly outperform others.
This review confirms that deep learning has demonstrated strong performance for breast mass classification and segmentation using ultrasound, often matching trained radiologists. However, the field is still overwhelmingly concentrated on B-mode imaging and on the diagnostic phase. The potential of other ultrasound modalities - particularly elastography and Doppler - combined with AI remains substantially underexplored.
Future research must expand into the prognostic domain: predicting whether cancer will recur, how patients will respond to chemotherapy, and what long-term outcomes patients can expect. These questions are clinically essential but have received far less attention than initial lesion classification in published deep learning literature.
There is a pressing need for prospective multicenter studies that collect data specifically for AI training and validation across diverse patient populations, scanner types, and geographic settings. Such studies would address the generalizability problem that limits most current models and build the evidence base needed for regulatory approval and clinical integration.
Integrating deep learning models with clinical metadata - such as patient age, hormonal status, family history, and prior imaging - could significantly improve predictive accuracy beyond what image analysis alone can achieve. The ultimate goal is a comprehensive, AI-augmented breast cancer management pathway from screening through treatment monitoring that reduces burden on radiologists and improves outcomes for patients worldwide.