The concept of AI was first introduced to the public by John McCarthy in 1956 during the Dartmouth Conference. AI aims to develop systems that assist with tasks requiring human expertise. In medical imaging, AI excels at analyzing complex image data to identify patterns and support diagnostic decisions, offering tools like radiomics for quantitative feature analysis (20).

AI infrastructure layer supported by machine learning (ML) core algorithms and AI infrastructure frameworks supports and realizes AI applications by learning the internal laws of data and training models (21). For medical image analysis, ML and its advanced subset, deep learning (DL), are the most relevant AI paradigms. Convolutional neural networks (CNNs), a specialized DL architecture, are particularly pivotal for ultrasound image interpretation. Their key characteristics are summarized in Table 1.

AI opens unprecedented possibilities for medical research, particularly in complex data analysis, pattern discovery, and predictive modeling. AI big data analytics are forming an emerging field through effective integration with medical imaging, especially noninvasive advanced imaging analysis, known as radiomics. AI analyzes lesion features extracted from large volumes of digital quantitative information from radiological imaging to provide predictive models. AI algorithms are trained by extracting and analyzing markers, such as pathology test results and genetic microcoding, from patient diagnosis or prognosis. The AI system processes image data from various imaging modalities to obtain the output information and continually undergoes self-correction and re-learning to make diagnostic and treatment decisions (22). Imaging physicians, after scrutinizing numerous image data for diagnosis, may suffer from fatigue, potentially resulting in missed diagnoses or misdiagnoses. AI has shown significant potential in revolutionizing medical diagnostics, offering high accuracy and efficiency in identifying diseases such as cancer at early stages and providing personalized treatment plans by analyzing vast amounts of medical data. to mimic the work of humans in medical image processing for a long period and make the correct diagnosis, improving the efficiency of radiologists by reducing interpretation time (23). AI-based radiomics is increasingly utilized in diagnostic ultrasound. The pre-training of AI models on extensive image datasets is a critical step in leveraging AI for radiomics, particularly in diagnostic ultrasound, as demonstrated by recent advancements in breast ultrasound and liver fibrosis staging. The pre-training of models extracts deep features from rich ultrasound images, which in turn can help identify new biomarkers or develop new criteria for diagnosing a specific disease. These trained models can assist clinicians in interpreting ultrasound images and automatically provide diagnostic results or recommendations, which not only saves time and improves access to healthcare but also optimizes ultrasound diagnosis.

In 2022, generative AI (GenAI) emerged as a transformative branch of AI, capable of creating new data instances that resemble the training distribution (24). Within breast ultrasound, GenAI holds immediate promise for addressing two pivotal challenges: image quality enhancement and data scarcity (25, 26). For image enhancement, GenAI models, particularly those based on Generative Adversarial Networks (GANs) or diffusion models, can be employed to reduce characteristic speckle noise, improve image resolution, and synthesize one imaging modality from another. These applications aim to improve lesion conspicuity and feature extraction, directly impacting diagnostic confidence. For data augmentation, GenAI can generate high-quality, annotated synthetic breast ultrasound images. This is crucial for expanding training datasets, balancing class distributions, and simulating rare or challenging presentations like non-mass lesions, thereby mitigating the data-hungry nature of deep learning models.

Currently, GenAI is increasingly being used in healthcare to inform clinical service functions. GenAI can improve diagnostic accuracy through knowledge acquisition for screening and diagnosing diseases. While GenAI is still in the process of being fully realized, the healthcare industry has already seen significant strides in its application, with 62% of healthcare and life science executives implementing GenAI solutions and 74% experiencing investment returns in at least one use case. Continued research and development are essential to deepen the integration of GenAI in healthcare, particularly in areas such as automated services, to further revolutionize drug development and personalized patient care.

However, a pronounced chasm separates this technical potential from validated clinical utility. The majority of GenAI applications in breast ultrasound are proof-of-concept demonstrations reliant on small, often idealized datasets, remaining far from clinical readiness. The most critical evidence gap, as highlighted earlier, is the near absence of robust, multi-center clinical trials. Such trials are essential to determine if GenAI-driven enhancements yield measurable, reproducible improvements in diagnostic accuracy and workflow efficiency across diverse clinical environments, or if reported gains are specific to constrained experimental settings.

The evolutionary path forward for AI in breast ultrasound points toward systems with enhanced multimodal integration and clinical context-aware decision support, moving beyond single-modality, task-specific classifiers. Rather than pursuing the broad and distant goal of Artificial General Intelligence (AGI), the near-term focus is on developing specialized AI systems that can emulate a radiologist’s integrative reasoning process. Such systems aim to automatically fuse and interpret diverse data streams available in a clinical setting. This includes correlating features across multiple ultrasound modalities from the same exam and incorporating relevant clinical context from electronic health records.

The technical foundation for this next stage lies in advancing multimodal learning architectures, cross-modal representation learning, and explainable AI (XAI) frameworks that make the integrated decision-making process transparent. By doing so, AI has the potential to evolve from an image analysis tool into a comprehensive clinical decision-support partner. It could provide a unified risk assessment, suggest differential diagnoses, and flag inconsistencies, thereby assisting in complex case evaluation and management planning.

This progression from an image analysis tool to a clinical decision-support system underscores the translational goal of AI: to augment and standardize the radiologist’s cognitive process, thereby enhancing the consistency and accuracy of diagnostic outcomes across diverse clinical settings and levels of expertise. Achieving this requires concerted efforts in creating large, curated, multimodal datasets and developing robust validation frameworks to ensure these advanced systems are safe, effective, and trustworthy in real-world clinical workflows.

ML is an emerging field that combines computer science and statistics. Unlike classical AI programming, which executes algorithms to produce results, ML enables computers to learn from input data without being explicitly programmed to use the dataset and associated outputs to generate algorithms that can describe the relationship between the two (27, 28).

DL cognitions features and patterns from data by constructing neural network models with multiple layers. It can autonomously “learn” from erroneous algorithmic outputs without human intervention, enhancing algorithmic stability and accuracy (34, 35). Compared to traditional ML methods, DL offers greater feature extraction capabilities and higher model complexity. That is capable of handling large-scale data and complex tasks.

In image recognition tasks, each input to an artificial neural network corresponds to a pixel, without considering the connectivity between nodes in the layer, potentially losing the spatial context of image features in practical applications (54).CNN, a key subset of DL, preserves the spatial relationships between image pixels through the combinations of input, convolutional, pooling, fully connected, and output layers to automatically learn features and patterns in images(see Figure 2) (32).

Radiomics combines large data analysis technology with medical imaging, which is mainly divided into five steps: medical image acquisition, ROI segmentation, feature extraction, feature selection, model building, and validation. Radiomics can acquire X-ray, US, CT, MRI, and even biopsy slice images. The regions related to the lesion in the image are summarized and segmented for further feature extraction. Next, radiomics like signal intensity, lesion shape, size, and texture features are extracted from ROI regions. Features with good repeatability, stability, and independence are selected by feature selection methods, such as minimum absolute contraction. Finally, model building and testing of independent samples were performed by common statistical methods and advanced ML strategies (64).

The clinical imperative of radiomics extends beyond feature extraction. Its core promise lies in translating complex quantitative data into objective, actionable biomarkers that can resolve diagnostic ambiguity in equivocal lesions, provide prognostic insights complementary to histopathology, and ultimately enable more personalized and confident patient management decisions.

Radiomics features mainly describe the heterogeneity of the internal structure of the lesions, and statistical features quantify the grayscale distribution and texture patterns of the images. By extracting these features, feature vectors can be constructed for subsequent ML algorithms for classification and diagnosis.

A critical challenge in radiomics is feature reproducibility. Extracted features are highly sensitive to variations in image acquisition parameters, pre-processing steps, and segmentation methods. Without rigorous standardization across centers, the generalizability of radiomics models is fundamentally limited. Many published studies fail to adequately address this issue or to perform validation on fully independent, external datasets, which is essential to assess true clinical applicability.

Consequently, the external validation of radiomics models on completely independent datasets, preferably from institutions using different equipment and protocols, is not merely a best practice but a prerequisite for assessing true clinical applicability. Many published high-performance models fail this critical test, their accuracy metrics reflecting optimistic bias inherent to single-center, retrospective studies. Rigorous external validation serves as the primary means to expose and quantify the impact of domain shift.

The development of radiomics has enabled the close integration of engineering intelligence with imaging medicine, presenting unprecedented opportunities for medical diagnosis and research. In disease diagnosis, radiomics can dig out hidden subtle information, objectively assess inter- and intra-tumor heterogeneity through spatial distribution, and assist physicians in more accurately determining the type, stage, and prognosis of diseases (65, 66). In treatment planning, radiation therapists determine the irradiation range and dose more accurately based on the detailed information extracted from images, optimize the automatic planning process and dosimetric trade-offs, and improve the therapeutic efficacy while reducing damage to the surrounding healthy tissues (67, 68). In disease monitoring and efficacy assessment, real-time tracking of disease progression and evaluation of treatment efficacy through image analysis allows for early detection of signs of recurrence or assessment of the effectiveness of treatment, leading to adjustment of treatment strategies.

The applications of radiomics in BC mainly include both lesion classification and treatment outcome prediction. Lesion classification involved differentiating benign and malignant lesions, molecular subtypes, and other clinicopathologic indices, including the status of the anterior sentinel lymph nodes. As a whole, radiomics is currently a popular direction expected to improve the accuracy of the diagnostic classification of breasts, with ultrasound-associated radiomics developing particularly rapidly (69).

Traditional ML methods rely on manual feature extraction and visual observation of image morphology by ultrasound workers, which requires domain knowledge and is difficult to capture complex patterns that may have significant inter-observer variations. The application of DL techniques in image classification, object detection, segmentation, and image synthesis provide a new perspective for the diagnosis of breast lesions. CNNs (e.g., ResNet, DenseNet), through the convolution layer CNNs (e.g., ResNet, DenseNet) automatically extract hierarchical image features. Their ability to characterize complex patterns, including subtle findings like architectural distortions and spiculated margins, is significantly superior to traditional methods. However, DL models rely on large-scale labeled data, and the high cost of medical image labeling has led to overfitting in some studies using small samples. The comparison of and advantages and limitations of ML and DL models in breast ultrasound is summarized in Table 3. In recent years, with the rapid growth in graphics processor computing power, more and more studies have integrated DL and radiomics (73–75). Approaches that combine DL and radiomics utilize supervised learning techniques. Employing learning methods to preprocess limited image data and derive numerous quantitative features from these images through the machine’s autonomous learning process. The subjectivity and uncertainty of radiologists’ classification choices are reduced to speed up the diagnosis and treatment and liberate the healthcare burden.

DL techniques are employed for fully automated breast density segmentation and classification, with results strongly correlating with radiologists’ manual classification. Zhang et al. aim to investigate whether DL-based radiomics models can improve the diagnostic performance of ultrasound for breast lesion classification. They developed DL radiomics models based on B-US and SWE in terms of breast lesion classification and compared them with quantitative SWE parameters and diagnostic assessments by radiologists. The study found that the area under the ROC curve for both the B-US and SWE-based DL radiomics models in the training cohort was 0.99 (95% CI: 0.99 - 1.00). Both DL radiomics models were more specific than the maximum elasticity parameter in both the training and independent validation cohorts, and the DL radiomics model significantly outperformed the quantitative SWE parameter and the BI-RADS assessment for breast lesion classification. Huang et al. found that the CNN-based tumor identification and grading system had an accuracy BI-RADS-3 of 0.99, BI-RADS-4a of 0.94, BI-RADS-4b of 0.73, BI-RADS-4c of 0.92 (76). Ciritsis et al. similarly classified benign and malignant breast ultrasound image classification by BI-RADS and CNN modeling. CNN modeling accuracy for BI-RADS was slightly higher than that of radiologists (ACC: 0.93 vs 0.92) (77). These studies suggest that the integration of DL-based radiomics methods can mimic the human decision-making process and help improve ultrasound’s ability to classify the benign and malignant nature of breast lesions.

However, these impressive accuracy metrics require cautious interpretation. The aforementioned studies are predominantly retrospective and single-center in design, which may introduce selection bias and limit generalizability. Performance achieved in optimized research environments may degrade significantly when models encounter data from different institutions, ultrasound machines, or patient populations—a key test of real-world utility that many models have not yet passed. Therefore, while DL models show great potential, their clinical maturity and readiness for deployment should be viewed as preliminary, pending validation through robust, prospective, multi-center studies.

While these reported accuracies are promising, they require careful interpretation. The performance is typically achieved in optimized, retrospective research environments using single-center data. A major and frequently under-reported limitation is the potential for significant performance degradation when models encounter data from different institutions, ultrasound machines, or patient populations: a key test of real-world utility that many models have not yet passed.

Consequently, for AI to earn a definitive role in clinical pathways, validation must evolve. Future studies should prioritize prospective, multi-center trials that report not only standard accuracy metrics but also clinically meaningful endpoints. These include the reduction in unnecessary biopsies or short-term follow-ups, changes in radiologists’ diagnostic confidence and inter-reader agreement when using AI assistance, and the system’s impact on overall workflow efficiency.

DL models are also more sensitive for identifying malignancy in non-mass breast lesions that do not strictly meet the BI-RADS definition. Non-mass breast lesions that lack distinctive margins, shape, and typical ultrasound features are more difficult to diagnose. These lesions have higher malignancy rates and poorer prognosis and quality of life for patients. Li et al. demonstrated the MobileNet model for the benign-malignant differentiation of non-mass breast lesions achieved an AUC of 0.84 (95% CI: 0.81 - 0.86) in the test set (78). Another study looked for imaging features associated with non-mass malignant breast lesions in the development dataset by multivariate LR and found that the development dataset in the ultrasound classification system had a higher AUC compared to when it was not applied (0.91 vs 0.95, P < 0.05) (79). The development of DL model imaging histology may improve the accuracy of early diagnosis of non-mass breast lesions, ultimately improving patient prognosis.

Furthermore, integrating clinical data and pathological outcomes with pre-treatment ultrasound images can enhance the accuracy of breast lesion classification. This is achieved by leveraging radiomics, which employs deep learning models for classification, in conjunction with pertinent clinical information.

Currently, AI models based on CNNs such as VGG, ResNet, and DenseNet have been applied to some medical image classification fields, improving the screening accuracy of lung nodules, breast cancer, and other diseases. However, the ethical and visualization problems associated with the “black box” nature of complex models such as deep learning remain to be solved. AI models also face significant generalization challenges in cross-institutional applications. Multicenter studies have shown that device differences, inconsistent annotations, and non-standardized acquisition can all lead to reduced diagnostic accuracy of AI models. Synergistic advancement through domain-adaptive algorithms, federated learning frameworks, and clinically standardized workflows is the key to moving AI from the lab to the clinic, where it is needed. XAI technologies such as Gradient-weighted Class Activation Mapping (Grad-CAM) are beginning to be integrated to enhance the transparency of how models are processed. XAI technologies enable clinicians and researchers to track which features in an ultrasound image affect the model’s predictions and diagnostic decisions, ensuring that decisions are based on relevant image lesion features rather than artifacts in the data. This improves the trust and clinical applicability of AI models. In addition, XAI can identify whether the model is relying on flaws such as incorrect lesion features, helping engineers to improve their algorithms in a timely manner (80).

The XAI technology system is mainly divided into intrinsic and extrinsic interpretability (81). Intrinsic interpretability is based on the structural characteristics of the model, such as the weight coefficients of linear regression, the splitting rule of decision tree, etc., which naturally possesses human-understandable decision logic. Extrinsic interpretability interprets black-box models by means of post-hoc analysis, which is mainly run by feature importance analysis, visualization techniques, counterfactual interpretation, etc. SHAP (SHapley Additive exPlanations) is mainly based on the principle of game theory, which quantifies global feature importance by quantifying the contribution of each feature to the model prediction. In interpreting breast cancer AI models, SHAP values can point to a max influence of “number of calcified foci” on malignancy judgments. LIME (Local Interpretable Model-agnostic Explanations) explains local decision logic by generating linear models near the point of prediction. Grad-CAM localizes the region of interest of the CNN model in the image through gradient backpropagation.

While the current use of XAI technology has made model predictions more transparent, helping clinicians validate and trust AI-driven diagnoses. However, the development of medical XAI still faces significant challenges. The difficulty in translating heat map model results to clinical imaging terminology still biases physicians’ trust and use of model results (82).

For AI models to realize their potential, diagnostic accuracy is a prerequisite, but their ultimate value will be determined by seamless integration into existing clinical workflows and their tangible impact on patient management decisions. Current evidence supporting such integration and impact remains limited. The implementation of AI in breast ultrasound can be envisioned in several roles: as a concurrent read assistant providing real-time feedback during image acquisition, a second reader for independent verification post-scan, or an automated triage tool flagging high-priority cases. Each mode presents distinct implications for workflow efficiency, radiologist responsibility, and required regulatory approval.

The most significant clinical impact of AI is anticipated in refining the management of diagnostically challenging cases, particularly within the BI-RADS 3 and BI-RADS 4a categories. By supplementing subjective visual assessment with quantitative risk scores, AI has the potential to reduce unnecessary short-term follow-ups for stable, very low-risk BI-RADS 3 lesions, and to decrease benign biopsy rates for BI-RADS 4a lesions by more confidently upgrading a high-risk subset. This optimization of the diagnostic pathway can alleviate patient anxiety, improve healthcare resource allocation, and potentially expedite treatment for aggressive malignancies. A forward-looking step would be the formal exploration of integrating AI-derived quantitative risk scores as a supplementary descriptor within the BI-RADS lexicon, potentially leading to a more nuanced and personalized risk assessment framework.

Successful translation, however, hinges on overcoming significant non-technical barriers that extend far beyond diagnostic accuracy. Clinician trust and acceptance must be cultivated, which depends not only on model interpretability but also on consistent performance within local clinical contexts and clear protocols for reconciling AI suggestions with clinical judgment. Seamless workflow integration requires adapting standardized reporting frameworks to incorporate AI outputs meaningfully, without creating ambiguity or additional burden. Furthermore, navigating the regulatory pathways for AI as a medical device demands a higher level of evidence, typically through prospective trials that demonstrate tangible clinical utility. Perhaps most critically, the widespread adoption of AI necessitates the development of clear medico-legal and ethical frameworks to address accountability, liability, and algorithmic bias, ensuring that these tools are deployed responsibly and equitably. Finally, demonstrating conclusive health economic value through tailored cost-effectiveness analyses is essential for sustainable implementation within diverse healthcare systems. Future development must, therefore, transition from proof-of-concept accuracy studies to pragmatic trials that measure clinically relevant endpoints such as the positive predictive value of AI-guided biopsy recommendations, changes in diagnostic confidence, and time-to-treatment intervals. The in-depth integration of XAI and medical knowledge still has a long way to go, and there is an urgent need to combine text, images, videos, and other forms to improve the intuitiveness of the explanation. Secondly, XAI technology has limitations: SHAP values have global interpretation consistency, but require thousands of model inferences and high computational complexity; LIME provides local interpretability, but may introduce bias due to simplified assumptions; Grad-CAM visualization is intuitive, but can only locate spatial regions and cannot quantify feature contribution. This difference in technical characteristics requires a multi-method joint interpretation strategy for clinical applications.

S-Detect is regarded as the most crucial image analysis software program for commercial CAD systems for the breast. S-Detect directly pinpoints breast lesions and manually selects ROI to detect lesions on the ultrasound devices even after freezing the image and uploading it to the workstation. The addition of S-Detect clearly distinguishes the features with high malignant representation by outlining the ROI independently and improving the diagnostic efficiency.

S-Detect improves the use of breast ultrasound in clinical practice and assists radiologists in making correct diagnostic decisions. Kiwook K. et al. conducted the first performance study of an AI-driven tool involving 192 breast lesions. They found that S-Detect classified diagnoses with a significantly higher AUC than radiologists (0.73 vs. 0.65, P = 0.04). This aligns with broader research indicating AI’s potential to enhance diagnostic accuracy in breast cancer. A study determining the effect of S-Detect on the diagnostic ability of radiologists of different experiences found that the diagnostic accuracy of inexperienced radiologists improved significantly with CAD assistance, as well as the SPE and PPV of experienced radiologists. Diagnostic results between radiologists and S-Detect showed moderate concordance (Kappa = 0.58) (87, 88). Wang et al. demonstrated significant improvement in AUC for radiologists of different experience when using S-Detect for categorical diagnosis (89). S-Detect shows potential to augment diagnostic accuracy, particularly for less-experienced readers. However, its utility is not uniform; performance is highly dependent on the chosen BI-RADS threshold and remains subject to variability from manual ROI selection. Most importantly, its effectiveness and consistency across diverse clinical settings lack robust validation through large-scale, multi-center prospective studies, which is necessary before broad clinical recommendations can be made.

S-Detect also improves inter- and intra-observer concordance (90). The final assessment of inter-observer variability by category for ultrasound was significantly improved (P < 0.001) in Park’s study (87).

Different definitions of benign and malignant for BI-RADS may lead to different diagnostic efficacy of S-Detect. The diagnostic efficacy of S-Detect was significantly higher than that of radiologists when the threshold of benignity and malignancy of the lesion was set to the BI-RADS-4a (ACC: 0.71 vs 0.56, P < 0.05). However, the diagnostic efficacy of S-Detect was significantly lower than that of radiologists (ACC: 0.71 vs 0.87, P < 0.05) when the threshold value was set to BI-RADS-4b (89). At the same time, S-Detect also leads to changes in diagnostic BI-RADS classification (91, 92). However, the diagnostic performance of S-Detect demonstrates significant dependency on BI-RADS threshold definitions. This performance discrepancy underscores potential limitations in risk stratification across BI-RADS categories. Additionally, the requirement for manual ROI selection introduces operator-dependent variability, as inter-operator differences in lesion contouring may influence feature extraction and classification outcomes. Multicenter validation studies are currently lacking to establish the generalizability of diagnostic consistency across diverse clinical settings. The analysis of the benignness or malignancy of BI-RADS lesions by S-Detect remains questionable, and S-Detect does not completely liberate radiologists from ultrasound diagnosis. The stability of the diagnostic efficacy of the S-Detect software for assisted classification of breast lesions under different definitions and its adaptability to the BI-RADS should be continuously explored in clinical practice.

In addition to S-Detect software, CAD systems such as MobileNet, KOIOS, and others have shown good value in breast lesion classification.

MobileNet can accurately classify breast lesions, especially BI-RADS-4 breast lesions (93). A screening study involving 479 BI-RADS-4a breast lesions found that the MobileNet model performed best in classification (AUC: 0.90, ACC: 0.91). It was predicted that 14.4% of BI-RADS-4a patients would be upgraded to BI-RADS-4b as a result of MobileNet modeling (94). The good performance of MobileNet for benign and malignant differentiation of non-mass breast lesions has also been described previously (78). Compared with the base MobileNet model, the novel MobileNet-v2 model improved based on super-resolution ultrasound images and demonstrated better diagnostic performance in the breast lesion classification, with its AUC improved by 0.09 and 0.03 in the training and test sets, respectively (95). The MobileNet model can reduce the incidence of malignant tumors in breast lesions diagnosed as BI-RADS-4a on preoperative ultrasound examination, enabling clinicians to strengthen the attention to such patients, make timely and correct management, and avoid delays. Although MobileNet achieves a great AUC for BI-RADS classification, but struggles with non-mass lesions. Transfer learning-based models like MobileNet-v2 require large-scale, standardized datasets to avoid overfitting.

KOIOS DS TM (KOIOS) has demonstrated excellent performance as a new CAD system in providing risk assessment for breast malignancy. KOIOS redefines breast lesion classification based primarily on the BI-RADS (see Table 4). KOIOS, like S-Detect, improves the correct assessment of ultrasound breast lesions for most physicians (95). Its decision support for the assessment of breast ultrasound lesions may reduce unnecessary biopsies. KOIOS has a significant impact on the assessment of lesions in the breast and on decisions about whether to biopsy a lesion (96). The SEN of KOIOS for performing breast lesion classification and deciding whether to biopsy a lesion was 0.87 (95% CI: 0.84 - 0.90), whereas the mean SEN for radiologists’ judgments from ultrasound images only was 0.83 (95% CI: 0.78 - 0.89) (97). Patients with KOIOS-recommended biopsies had a higher rate of positivity compared with BI-RADS-recommended biopsies (98). At the same time, KOIOS may also improve inter- and intra-observer agreements. However, it has been criticized for underperforming in patients with dense breast tissue (99). Its reliance on multimodal radiomics may introduce bias when clinical data is incomplete (100).

Despite the vigorous development and achievements of S-Detect and KOIOS software, CAD systems for BC screening continue to face numerous challenges, notably the absence of a global public dataset, complexities in binary classification, inadequate image quality, and a heavy reliance on manual ROI annotation. There is still a long way to go in the development of commercial CAD systems.

Studies developing novel AI models (e.g., customized CNNs, GANs for enhancement) predominantly follow a single-center, retrospective design. While these works demonstrate impressive accuracy (often AUC >0.90) and drive methodological innovation, their results are intrinsically tied to the local data domain. Performance claims are vulnerable to domain shift and may not hold across different ultrasound machines, patient demographics, or institutional practices. The primary value of this paradigm lies in proof-of-concept and algorithm advancement, not in proving clinical readiness.

Commercial systems like S-Detect and KOIOS DS represent a later stage in the translational pipeline. Their development increasingly incorporates data from multiple sites, albeit often under controlled conditions. More importantly, their evaluation increasingly includes multi-center validation studies (e.g., references 87, 97). While reported diagnostic metrics may sometimes appear more modest than research prototypes, they often reflect performance across a broader range of clinical scenarios. The focus shifts towards workflow integration, user interface design, regulatory clearance, and demonstrating consistent utility across diverse settings—key factors for clinical adoption that are frequently absent in research papers.

The most significant distinction between a research prototype and a clinically viable tool is the evidence from prospective, externally validated trials. Future research must prioritize this step. Furthermore, explicit strategies to mitigate domain shift are needed. These include: (1) Developing and reporting model performance stratified by ultrasound device vendor or model; (2) Utilizing domain adaptation techniques in model training to improve robustness to acquisition differences; (3) Federated learning frameworks that enable training on distributed data without centralizing sensitive patient information, thereby inherently encompassing multi-center heterogeneity.

The creation of large-scale, publicly available datasets, acquired from diverse ultrasound systems and adhering to standardized imaging protocols, is a fundamental prerequisite. These should serve not only for training but as independent benchmark platforms to rigorously evaluate model generalizability and resilience to domain shift:a key limitation highlighted in current research.

Future validation must transition from retrospective accuracy studies to prospective, multi-center trials designed with regulatory endpoints for Software as a Medical Device. These trials should demonstrate improvement in tangible clinical outcomes, such as increased positive predictive value of biopsies or reduced time to definitive diagnosis, providing the evidence base required for FDA, CE, or NMPA clearance.

For seamless adoption, AI’s role must be defined within the radiological workflow. Its output requires structured integration into the BI-RADS reporting framework. This could involve providing a quantitative malignancy risk score alongside the BI-RADS category or offering decision-support prompts to aid in the management of equivocal lesions, thereby directly influencing patient management algorithms.

Beyond technical and clinical validation, the responsible integration of AI into healthcare demands proactive engagement with its broader regulatory, ethical, and legal ecosystem. Future development must align with evolving standards for Software as a Medical Device, requiring early dialogue with regulatory bodies and study designs that meet approval requirements. Concurrently, the field must establish ethical guidelines and governance structures to ensure fairness, transparency, and accountability in algorithm development and deployment. Clear medico-legal protocols are also needed to define the standard of care and shared responsibility in AI-assisted diagnosis, addressing critical questions of liability and data privacy. Navigating these complex dimensions is not ancillary but fundamental to building the trust required for AI to become a viable and sustained component of clinical practice.

Beyond technical and clinical validation, successful deployment necessitates implementation science studies to identify barriers to adoption and robust health economic analyses to demonstrate cost-effectiveness within specific healthcare systems. This evidence is crucial for stakeholder buy-in and sustainable integration.

In conclusion, the journey from promising research to routine practice hinges on concerted efforts across multiple interconnected domains: the creation of standardized, multi-vendor benchmarks; the generation of prospective, regulatory-grade clinical evidence; the development of frameworks for structured clinical integration; the proactive navigation of regulatory, ethical, and legal landscapes; and the demonstration of practical implementation value through health economics research. By prioritizing these comprehensive translational imperatives, the field can ensure that AI evolves into a reliable, trusted, and equitably deployed standard of care in breast ultrasound.