Breast cancer generates approximately 3,000 new research publications every month, creating an information volume that exceeds the capacity of individual clinicians to fully integrate. Artificial intelligence offers a path to managing this complexity - helping synthesize large datasets, support clinical decisions, and personalize treatment in ways that individual human review alone cannot reliably achieve at scale.
This narrative review from an international team including researchers from Dana-Farber Cancer Institute, the European Institute of Oncology, and the German Breast Group surveys the current state of AI applications across all major treatment domains in breast cancer: surgery, radiation therapy, systemic treatment, and supportive care. Rather than providing an exhaustive catalog, the review focuses on the most clinically relevant and ready-to-implement applications.
A central conclusion is that AI's readiness for clinical use is highly uneven across breast cancer treatment domains. Some applications - particularly organ-at-risk autocontouring in radiation oncology and certain preoperative decision support tools - are already in routine clinical use. Others, including AI-assisted systemic treatment decision-making and patient supportive care, remain too immature for reliable clinical deployment.
The review explicitly addresses not just what AI can do but how clinicians should approach AI responsibly - including detailed discussion of the biases that can undermine model reliability, the distinction between FDA-cleared and rigorously FDA-approved tools, and the organizational requirements for safe, sustained AI deployment in oncology practice.
In the preoperative setting, AI has been applied to predict pathological complete response (pCR) after neoadjuvant chemotherapy, with the goal of potentially omitting surgery for patients who achieve a complete response. Current results show AI can identify true non-responders with near-zero false negative rates, but specificity remains limited at around 65%, meaning many patients without a true pCR would be incorrectly classified as responders - an unacceptable safety margin for surgical omission decisions.
AI tools for axillary lymph node staging - assessing whether cancer has spread to underarm lymph nodes without requiring invasive surgical sampling - have shown more promise. The NILS web-based neural network achieved 90% sensitivity for predicting lymph node involvement from clinical parameters alone, and a combined AI-plus-clinician approach for PET/CT-based assessment improved sensitivity by 10% over clinician review alone. AI-assisted reconstruction planning tools can identify vessel perforators for flap surgery more accurately than radiologist teams while saving 2-3 hours of radiologist time.
Intraoperative AI for surgical margin assessment - determining whether enough normal tissue surrounds the tumor during breast-conserving surgery to reduce re-excision rates - has achieved sensitivity around 84-96% and specificity of 42-76% across different methods. Novel techniques using ultraviolet fluorescence scanning microscopy with deep learning reached 100% sensitivity and 87.5% specificity for margin detection, but require specialized equipment not currently available in most operating rooms.
Postoperative AI has shown utility in predicting complications from reconstructive surgery. Neural network models predicted risk factors for abdominal flap reconstruction complications with 81% accuracy and 89% specificity, and machine learning models predicted periprosthetic infection risk with up to 82% accuracy - potentially identifying high-risk patients who would benefit from enhanced prophylaxis or monitoring before complications arise.
Radiation oncology represents the breast cancer treatment domain where AI is most clinically established. Deep learning-based organ-at-risk (OAR) autocontouring - automatically delineating the heart, lungs, esophagus, and contralateral breast that must be protected from radiation - can be completed in as little as 10 seconds with AI assistance and requires major corrections in only 2% of cases. This compares to manual contouring that typically takes 20-30 minutes per patient.
Beyond efficiency, AI autocontouring meaningfully improves clinical consistency. In a 31-center multicenter evaluation, AI assistance reduced inter-expert variation for OAR contouring and improved inter-expert agreement for target volume delineation by up to 19%. A separate study found that guideline consistency in target delineation improved from 77.1% to 90.7% with AI assistance - a clinically significant improvement with direct implications for treatment quality and safety across different centers and individual clinicians.
AI models for predicting cardiac dose from chest X-rays can identify patients likely to benefit from Deep Inspiration Breath Hold (DIBH) radiotherapy - a technique in which patients hold their breath to displace the heart away from the radiation field. CNN-based models achieved AUC 0.86 for predicting which patients would achieve more than 1 Gy mean heart dose reduction with DIBH, enabling objective pre-planning selection of this specialized delivery technique.
Automated treatment planning using AI has produced plans that in blinded review by expert radiation oncologists were rated acceptable at 95% - outperforming physician-generated plans at 90% in one head-to-head study. AI also enables surface-guided radiotherapy positioning to be automated from a 120-second manual process to 1.2 seconds, and enables online-adaptive radiotherapy workflows where treatment plans are automatically adjusted at each daily session based on real-time imaging of anatomical changes.
AI decision support tools for systemic breast cancer treatment range from guideline-based recommendation systems to large language models (LLMs). The most studied tool, IBM Watson for Oncology (WFO), showed concordance with multidisciplinary tumor board (MTB) decisions in 80-93% of cases depending on disease stage and institutional setting. However, concordance dropped substantially for advanced-stage disease and older patients - precisely the cases where reliable decision support would be most valuable.
Large language models including ChatGPT have been evaluated for both tumor board support and patient-facing information. ChatGPT 3.5 achieved only 16-58% concordance with MTB decisions depending on the evaluation, made significant clinical errors including incorrectly recommending targeted therapy for HER2-negative patients, and cannot maintain up-to-date guideline knowledge due to training cutoff dates. Current commercially available LLMs are therefore considered not ready for clinical use in treatment decision support.
AI models predicting response to neoadjuvant chemotherapy - identifying in advance which patients will achieve pCR - have shown more encouraging results. A deep convolutional neural network analyzing digitized pathology slides achieved AUC 0.751 overall and 0.88 for stage I disease. A multi-omic machine learning model combining tumor mutation profiles, copy number alterations, immune infiltration, and T cell markers achieved AUC 0.87 - demonstrating that integrating biological data types meaningfully outperforms models using any single data source.
Despite these results, the review cautions that outcome prediction studies are frequently affected by methodological biases. A metanalysis identified widespread problems with reporting standards, dataset selection, and overfitting in AI treatment outcome prediction research. The caution against overinterpreting small single-institution AI studies applies particularly strongly to treatment decision support, where errors in recommendations could directly harm patients.
A critical practical concern for AI in breast cancer care is the variety of biases that can enter models at different stages of development. Historical bias occurs when training data reflect longstanding healthcare disparities - for example, models trained predominantly on data from academic centers may perform poorly for patients treated in community settings. Representation bias arises when training populations do not reflect the diversity of patients the model will be applied to, leading to differential performance across demographic groups.
Sampling bias, omitted variable bias, measurement bias, and labeling bias each represent distinct pathways through which inaccurate or skewed data can corrupt model performance in ways that may not be apparent during development but manifest during deployment. Omitted variable bias is particularly insidious in oncology, where excluding prior treatment history, comorbidities, or relevant biomarker data can substantially impair a model's predictive validity.
Most AI tools in clinical practice are regulated differently from drugs. In the United States, most AI algorithms are FDA-cleared under Section 510(k) - a substantial equivalence pathway - rather than formally approved through the more rigorous premarket approval process. Clinicians using these tools may lack access to information about the training dataset, validation approach, or patient populations used to develop the model - making independent evaluation of reliability difficult.
Addressing bias requires strategies across the full model lifecycle. During development: multiple imputation for missing data, rebalancing skewed datasets, and fairness-aware regularization. During deployment: multidisciplinary teams combining oncologists, data scientists, and social scientists to ensure clinical and ethical oversight. After deployment: continuous performance monitoring, because model performance can drift over time as patient populations, imaging equipment, and clinical protocols evolve - a phenomenon called dataset shift.
Federated learning is a training approach that enables AI models to learn from data distributed across multiple institutions without sharing raw patient records. Each institution trains a local model on its own data, and only model updates (not patient data) are exchanged and aggregated into a shared global model. This preserves privacy while enabling models to learn from the diverse patient populations and imaging protocols present across different hospitals and geographic regions.
Federated learning is particularly relevant for addressing geographical and infrastructural biases in AI performance. Models trained exclusively at large academic centers in high-income countries may fail to generalize to community hospitals, lower-resource settings, or patient populations with different demographics and disease presentations. Federated training across geographically and institutionally diverse datasets can produce more robust and equitable models.
The potential of AI to reduce healthcare disparities depends on deliberate effort to include underrepresented populations in model development. AI models that perform poorly for certain demographic groups could amplify existing care inequities rather than reduce them - making diversity in training data a matter of both scientific validity and health equity. The review emphasizes that equitable AI is not simply desirable but essential to the ethical deployment of these tools in cancer care.
Implementing robust AI programs in clinical settings also requires hybrid teams combining clinicians, data scientists, and engineers with defined responsibilities for ongoing model monitoring, recalibration when performance degrades, and communication with clinical users about AI limitations and appropriate use. Treating AI deployment as a one-time implementation rather than an ongoing operational responsibility is a common failure mode that can lead to patient harm as model performance drifts over time.
The review's overall assessment is that AI in breast cancer treatment has achieved clinical readiness in a limited but important set of applications. In radiation oncology, OAR and target volume autocontouring is already in routine clinical use, providing time savings and consistency improvements that are well validated. Online-adaptive radiotherapy workflows incorporating AI are clinically operational at leading centers, and AI-assisted treatment planning is approaching clinical-grade quality.
Several near-term opportunities are identified. Automated radiotherapy treatment planning, image-guided delivery optimization, and outcome prediction after neoadjuvant treatment represent areas where AI performance is strong enough to anticipate near-term clinical deployment with appropriate validation. In surgery, AI assistance for breast reconstruction planning and postoperative complication prediction are the most promising near-term applications.
Areas where AI remains far from clinical usability include AI-assisted systemic treatment decision-making, support for surgical de-escalation decisions (omitting breast or axillary surgery after NAC), and patient supportive care applications. Current generative AI tools such as ChatGPT have demonstrated insufficient concordance with clinical expert decisions and unacceptable error rates for routine use in treatment planning discussions.
Prospective validation and post-deployment performance monitoring are highlighted as the most critical unmet needs for the field. Most current evidence comes from retrospective single-institution studies - a design that is well-suited for discovering AI capabilities but insufficient to establish clinical safety and generalizability. Moving AI from research to standard clinical practice in breast cancer will require the same rigorous evaluation applied to therapeutic interventions: prospective trials, independent validation, and systematic surveillance of real-world performance.