In a significant stride towards revolutionizing breast cancer diagnostics, researchers Huang Yingmin and Chen Zhiyi from the Department of Ultrasound Medicine at The Third Affiliated Hospital of Guangzhou Medical University have published a comprehensive review detailing the transformative potential and existing hurdles of Artificial Intelligence (AI) in breast ultrasound imaging. Their work, appearing in the February 2021 issue of Guangdong Medical Journal, synthesizes the latest advancements and offers a critical perspective on the future integration of AI into clinical workflows. This is not merely an incremental improvement; it represents a fundamental shift in how medical professionals approach the detection, classification, and management of breast disease, promising a future of enhanced precision, efficiency, and personalized care.
The global burden of breast cancer is immense and growing. As the most common cause of cancer-related death among women worldwide, early and accurate diagnosis is paramount for improving survival rates and patient outcomes. Conventional breast ultrasound, while a cornerstone of screening and diagnostic protocols due to its accessibility, safety, and real-time capabilities, is inherently limited by its dependence on the operator’s skill and experience. This subjectivity can lead to variability in interpretation, missed diagnoses, and unnecessary biopsies, creating a critical need for more objective, standardized, and reliable tools. Enter Artificial Intelligence. AI, particularly through its subfields of machine learning and deep learning, offers a powerful solution to these longstanding challenges. By training algorithms on vast datasets of annotated ultrasound images, AI systems can learn to recognize subtle patterns and features that may be imperceptible or inconsistently interpreted by the human eye. The goal is not to replace radiologists but to augment their expertise, acting as a tireless, highly consistent second reader that can flag potential malignancies, quantify tumor characteristics, and provide decision support, thereby reducing diagnostic errors and improving overall efficiency.
One of the most immediate and impactful applications of AI in this domain is Computer-Aided Detection (CADe). The primary function of CADe is to act as a safety net, meticulously scanning ultrasound images to identify and highlight suspicious regions that a human observer might overlook. This is especially crucial in dense breast tissue, where tumors can be particularly difficult to visualize. Early CADe systems for breast ultrasound were often semi-automated, requiring the sonographer to manually define a region of interest, which could inadvertently limit the system’s scope and effectiveness. The advent of Automated Breast Ultrasound (ABUS) has been a game-changer. ABUS systems perform a standardized, hands-free volumetric scan of the entire breast, generating a comprehensive set of images for review. This standardization provides the perfect foundation for fully automated CADe. Systems like the QView Medical platform, the first ABUS-CAD to receive FDA approval, can automatically detect lesions as small as 5 millimeters and provide navigational markers for potential malignancies. Clinical studies have demonstrated that when radiologists use ABUS-CAD, their sensitivity in detecting cancer improves significantly. However, this heightened sensitivity often comes at the cost of reduced specificity, meaning more benign lesions are flagged as suspicious, leading to an increase in false positives. This trade-off is a key area of ongoing research. More advanced systems, such as the one developed by Moon and colleagues using a 3D convolutional neural network, have shown remarkable promise in mitigating this issue. Their model achieved a sensitivity of 95.3% while simultaneously reducing false positives by 56.8%, a substantial improvement that could make screening more efficient and less anxiety-inducing for patients. Despite these advances, challenges remain. Current ABUS-CAD systems can struggle with very small tumors or those that do not exhibit the classic hypoechoic (dark) appearance, such as some isoechoic or hyperechoic lesions. Furthermore, the complexity of analyzing data across three dimensions can lead to “overfitting,” where the algorithm becomes too specialized to the training data and loses its ability to generalize to new, unseen cases, highlighting the need for more robust and adaptable models.
Beyond detection, the accurate classification of breast lesions as benign or malignant is equally critical. This is the domain of Computer-Aided Diagnosis (CADx). The visual characteristics of benign and malignant tumors on ultrasound often overlap, making differentiation a complex task even for experienced radiologists. AI-powered CADx systems address this by quantifying tumor features and using them to generate a diagnostic probability. There are two main approaches to CADx. The first is based on “hand-crafted” features, where experts manually define specific characteristics—such as texture, shape, and margin irregularity—for the algorithm to analyze. For instance, texture analysis has been shown to be effective in distinguishing the highly aggressive triple-negative breast cancer from benign fibroadenomas, which can sometimes appear similar. Combining texture with morphological features has further been shown to enhance diagnostic accuracy. One study using this approach on ultrasound images achieved an impressive accuracy of 88%, with a sensitivity of 81% and a specificity of 91%. The integration of the standardized BI-RADS lexicon into AI systems also provides a valuable framework. AI can be trained to interpret and apply BI-RADS categories more consistently, reducing inter-observer variability and improving the structure and clarity of radiology reports. For example, AI can help sub-categorize BI-RADS 4 lesions, which are considered suspicious but not definitively malignant, into more refined risk levels, aiding in better patient management. However, the limitation of this approach is its reliance on human expertise to define the features. It is labor-intensive, potentially subjective, and may miss subtle, complex patterns that are not easily codified by human experts.
This is where Deep Learning (DL), particularly Convolutional Neural Networks (CNNs), represents a paradigm shift. Unlike traditional CADx, DL algorithms do not require humans to pre-define features. Instead, they learn directly from the raw pixel data of the ultrasound images, automatically discovering the most relevant and discriminative patterns for diagnosis. This ability to uncover hidden, high-level features has led to DL models that rival, and in some cases surpass, human performance. Studies from Fujita, Tanaka, and others have developed CNN-based CADx systems for breast ultrasound that report diagnostic accuracies, sensitivities, and specificities all exceeding 90%. These systems can operate in real-time during an ultrasound exam, providing immediate feedback to the clinician and significantly boosting their diagnostic confidence and efficiency. The potential of DL extends even further into the realm of histological classification. Hizukuri and Nakayama developed a CNN model capable of distinguishing between four specific types of breast lesions: invasive carcinoma, non-invasive carcinoma, cysts, and fibroadenomas, with an accuracy ranging from 83.9% to 87.6%. This capability to predict the underlying tissue type non-invasively is a groundbreaking development that could streamline the diagnostic pathway and reduce the need for unnecessary biopsies. Despite these remarkable achievements, the “black box” nature of deep learning remains a significant barrier to its widespread clinical adoption. The lack of transparency in how these complex models arrive at their decisions makes it difficult for clinicians to trust and understand their outputs. For AI to be truly integrated into clinical practice, developing “explainable AI” (XAI) methods that can provide clear, interpretable rationales for their predictions is an urgent and active area of research.
Perhaps the most exciting frontier in AI-driven breast imaging is Radiomics. Radiomics moves beyond simple detection and classification to extract a vast array of quantitative features—hundreds or even thousands—from medical images. These features, which describe the tumor’s shape, texture, intensity, and heterogeneity, are then analyzed using machine learning to uncover patterns that correlate with underlying biological processes, genetic mutations, and clinical outcomes. While radiomics has been predominantly explored with MRI, recent research is unlocking its immense potential with ultrasound. The work of Li and colleagues demonstrated that a radiomics approach using multimodal ultrasound data could achieve an accuracy of 84.12% and a sensitivity of 92.86% in diagnosing breast cancer, showcasing its power for precision diagnosis. More profoundly, radiomics is beginning to bridge the gap between imaging and molecular biology. Breast cancer is not a single disease but a collection of molecular subtypes—such as Luminal A, Luminal B, HER2-positive, and Triple-negative—each with distinct prognoses and treatment responses. Traditionally, determining a tumor’s subtype requires an invasive tissue biopsy. Radiomics, however, offers a non-invasive alternative. Zhang and his team identified specific ultrasound features that correlate with different molecular subtypes. For example, Luminal A tumors were associated with an echogenic halo and posterior shadowing, while HER2-positive tumors were linked to posterior enhancement, calcifications, and rich blood supply. Similarly, Guo’s research found correlations between ultrasound features and the tumor’s receptor status and grade. Low-grade, ER/PR-positive cancers tended to be irregular and hypoechoic with posterior shadowing, while high-grade, triple-negative cancers were more likely to be regular in shape with posterior enhancement. This ability to non-invasively predict a tumor’s molecular profile and aggressiveness is revolutionary. It could allow clinicians to tailor treatment plans from the very beginning, selecting the most effective therapies for each patient’s specific cancer biology. Radiomics also shows great promise in predicting treatment response. For patients undergoing neoadjuvant chemotherapy (NAC) before surgery, predicting who will achieve a complete pathological response is crucial. Sannachi’s research used a radiomics model combining quantitative ultrasound and textural features to successfully predict NAC response, allowing for early adjustment of treatment for non-responders. Furthermore, radiomics can predict the status of axillary lymph nodes, a critical factor in staging and treatment planning. Zheng’s deep learning radiomics model, using features from conventional ultrasound and shear-wave elastography, was able to predict lymph node metastasis in early-stage breast cancer with high accuracy, potentially reducing the need for invasive sentinel lymph node biopsies. Despite its immense potential, radiomics faces significant challenges. The field lacks standardized protocols for feature extraction and validation. Different research groups use different methods, making it difficult to compare results and build upon previous work. Most studies to date are retrospective and involve relatively small patient cohorts. To realize its full clinical potential, large-scale, prospective, multi-center trials are urgently needed to validate these promising findings.
To gain an even more comprehensive understanding of a tumor, the field is moving towards Multi-modal Image Fusion. This technology intelligently combines information from different imaging modalities—such as ultrasound, MRI, CT, and PET—into a single, unified view. Each modality provides unique information: ultrasound offers real-time, high-resolution anatomical detail, MRI excels in soft-tissue contrast and functional imaging, and PET reveals metabolic activity. By fusing these datasets, clinicians can obtain a far richer and more accurate picture of the disease than any single modality could provide. A prime example is Ultrasound Volume Navigation, which overlays pre-acquired MRI or CT images onto real-time ultrasound. This is particularly valuable in pre-operative planning for breast cancer. MRI is highly sensitive and can detect additional lesions that are not visible on initial ultrasound. Volume Navigation allows the sonographer to precisely locate these MRI-detected, ultrasound-occult lesions during a follow-up scan, enabling accurate biopsy guidance and ensuring that the surgical plan accounts for all disease foci. Studies by Park and Aribal have confirmed that this technique is as effective as a dedicated second-look MRI, providing a more accessible and cost-effective solution. It significantly increases the detection rate of occult lesions and allows for safe and accurate biopsy regardless of the lesion’s depth. However, the technical challenges of multi-modal fusion are substantial. The biggest hurdle is image registration—precisely aligning images from different machines that have different resolutions, contrasts, and were taken at different times. Physiological changes in the patient between scans can further complicate this alignment. There is also a lack of standardized fusion algorithms and objective metrics for evaluating the quality of the fused image. The inherent noise and lower resolution of ultrasound images make them particularly difficult to register accurately with higher-resolution modalities like MRI. Despite these challenges, multi-modal fusion is widely seen as the future of AI in medical imaging, as it leverages the complementary strengths of different technologies to provide the most complete diagnostic information possible.
While the potential of AI in breast ultrasound is dazzling, Huang Yingmin and Chen Zhiyi offer a sobering and necessary perspective on the significant challenges that must be overcome before these technologies can be seamlessly integrated into routine clinical practice. The first and most fundamental challenge is data. Building robust, generalizable AI models requires massive, high-quality, and meticulously annotated datasets. In the medical field, such data is often siloed within individual hospitals, difficult to share due to privacy concerns, and plagued by inconsistencies in acquisition protocols, labeling errors, and missing information. Training an AI on noisy, biased, or incomplete data will inevitably lead to unreliable and potentially harmful outputs. There is a critical need for standardized data collection protocols and the creation of large, anonymized, publicly accessible databases to fuel the next generation of AI research.
The second major challenge is explainability. As mentioned earlier, the “black box” problem is a significant barrier to trust and adoption. If a radiologist cannot understand why an AI system flagged a lesion as malignant, they are unlikely to act on that recommendation, especially if it contradicts their own assessment. Developing AI systems that are not just accurate but also transparent and interpretable is essential. Researchers must create methods that can visually highlight the image features the AI used to make its decision, providing a clear and logical rationale that clinicians can understand and verify.
Finally, the authors raise profound ethical, legal, and regulatory questions that society must grapple with. Who is responsible if an AI system makes a diagnostic error? Should radiologists be required to undergo specific training before using AI tools? How should AI algorithms be validated and regulated to ensure their safety and efficacy across different patient populations and imaging equipment? There is a pressing need to establish clear ethical guidelines, legal frameworks, and standardized testing protocols to govern the use of AI in healthcare. These frameworks must prioritize patient safety, algorithmic transparency, and the protection of human rights.
In conclusion, the integration of AI into breast ultrasound is not a question of “if” but “when” and “how.” The work of Huang Yingmin and Chen Zhiyi paints a picture of a future where AI acts as a powerful, intelligent partner to the clinician, enhancing human expertise rather than replacing it. From improving the sensitivity of cancer detection with CADe to enabling non-invasive molecular subtyping with radiomics, the potential benefits for patients are enormous: earlier diagnoses, more accurate prognoses, and truly personalized treatment plans. However, this future is not guaranteed. It requires a concerted, multidisciplinary effort to address the formidable challenges of data standardization, algorithmic transparency, and ethical governance. Only by navigating these challenges with rigor and responsibility can we fully realize the promise of AI to transform breast cancer care from a reactive, one-size-fits-all model to a proactive, precise, and deeply personalized science. The journey has just begun, but the destination—a future where no breast cancer is missed and every treatment is perfectly tailored—is one worth striving for.
By Huang Yingmin, Chen Zhiyi, Department of Ultrasound Medicine, The Third Affiliated Hospital of Guangzhou Medical University. Published in Guangdong Medical Journal, Vol. 42, No. 2, February 2021. DOI: 10.13820/j.cnki.gdyx.20200788.