Breast cancer is the most common malignancy in women worldwide and the second leading cause of cancer-related death globally, with more than 600,000 deaths recorded in 2022 alone, predominantly in developing countries.
A critical problem is late diagnosis: roughly 30% of patients are already at an advanced stage when first detected, largely because early-stage disease produces few symptoms and current screening tools have significant limitations, particularly in women with dense breast tissue.
Breast cancer also exhibits profound molecular and immune heterogeneity, spanning hormone receptor status (ER/PR), HER2 expression, PAM50 molecular classes, and the tumor immune microenvironment (TIME). This heterogeneity drives highly variable responses to chemotherapy, targeted agents, and immunotherapy.
The 5-year survival rate for patients diagnosed at an early stage approaches 90%, but drops sharply at advanced stages, underscoring the urgent need for more accurate early detection and better tools to match patients with effective treatments.
Radiomics is the process of extracting large numbers of quantitative features from medical images -- such as shape, texture, and intensity -- and using them as measurable biological markers. When combined with AI, this transforms routine breast imaging into reproducible decision-support tools.
The AI workflow follows a structured pipeline: standardized multi-modal image acquisition (mammography, ultrasound, MRI, PET/CT), rigorous preprocessing, tumor segmentation, feature extraction, model training, and validation. Community guidelines from the Imaging Biomarker Standardization Initiative (IBSI) define consensus standards for 169 features, enabling results to be compared across different software and institutions.
Segmentation -- delineating exactly where a tumor begins and ends -- is foundational to the entire approach. Modern methods use deep learning architectures such as U-Net to automate this step, while also capturing information from the tissue surrounding the tumor, known as the peritumoral region, which often reveals important biological signals.
Model validation must go beyond single-center studies. Best practice includes nested cross-validation, testing on external multi-center datasets, calibration assessment, and decision-curve analysis to demonstrate that a model genuinely improves clinical decisions rather than just scoring well on internal metrics.
Multiple deep learning architectures have demonstrated impressive accuracy in distinguishing benign from malignant breast lesions. Combining the U-Net segmentation model with case-based reasoning achieved 91.34% diagnostic accuracy on mammography, while hybrid CNN architectures using ResNet50 and AlexNet with multi-class support vector machines reached 99% accuracy on curated benchmark datasets.
A large prospective Korean study (AI-STREAM) enrolling 24,543 women showed that AI-assisted reading increased cancer detection rate by 13.8% for breast-imaging subspecialists without raising recall rates, and specifically improved detection of small tumors under 20mm and early-stage, node-negative disease.
AI can also predict molecular subtypes non-invasively. DCE-MRI-based models predict ER, PR, and HER2 receptor status with AUC values reaching 0.89 for ER status. A model based on contrast-enhanced mammography features correctly distinguished Luminal from non-Luminal subtypes with 94% accuracy, while PET/CT models predicted HER2 status and molecular subtype with an AUC of 0.95.
A multi-instance learning framework (BBMIL) demonstrated that standard H&E-stained pathology slides alone contain enough information to predict ER/PR/HER2 status, PAM50 subtypes, and immune signatures, achieving AUCs of 0.883, 0.788, and 0.703 for ER, PR, and HER2 respectively -- making molecular profiling potentially accessible without costly genomic tests.
Axillary lymph node (ALN) status is one of the most important factors in breast cancer staging and treatment planning. Approximately 30 to 40% of early-stage patients have lymph node metastasis, and missed micrometastases can significantly affect survival outcomes.
Deep learning models applied to ultrasound images have substantially improved lymph node assessment. A Kohonen self-organizing model achieved 98% sensitivity and 99% specificity across 908 images, with an AUC of 0.97 -- significantly outperforming traditional feed-forward neural networks.
Multimodal approaches provide additional gains. Combining conventional ultrasound with shear-wave elastography (SWE) in a deep learning model predicted lymph node metastasis status with an AUC of 0.902 and metastatic load with an AUC of 0.905, both far exceeding single-modality performance.
MRI-based models have also shown strong results: a CNN applied to multiparametric MRI achieved 92.1% sensitivity (AUC 0.91) for detecting axillary metastasis, significantly outperforming radiologists (AUC 0.80). These non-invasive assessments could potentially reduce the need for surgical axillary dissection, sparing patients from complications like lymphedema.
Triple-negative breast cancer (TNBC) is particularly difficult to treat due to its high heterogeneity and lack of hormone or HER2 targets. AI-based approaches are now providing meaningful prognostic information for this challenging subtype by analyzing immune cell infiltration patterns and tumor microenvironment features.
A multi-omics signature (MLIIC) built from 25 machine learning algorithms and trained on immune-infiltrating cell characteristics significantly correlated with TNBC survival outcomes across multiple independent cohorts. Digitized spatial tumor microenvironment scores derived from deep learning analysis of tissue images achieved C-indexes of 0.65 and 0.60, outperforming the traditional TNM staging system for predicting overall and recurrence-free survival.
Predicting pathological complete response (pCR) -- the absence of remaining cancer after neoadjuvant chemotherapy -- is a key clinical goal. A fully automated multimodal pipeline (MIFAPS) integrating pre-treatment MRI, whole-slide pathology images, and clinical variables achieved AUC = 0.882 on external test sets and 0.909 on a prospective cohort of 1,004 patients, substantially outperforming any single-modality model.
A longitudinal model combining MRI habitat radiomics, bulk transcriptomics, and single-cell RNA sequencing data across 2,279 patients from 12 centers yielded AUCs of 0.863 to 0.888 for pCR prediction and revealed a biological link between imaging-derived phenotypes and B-cell infiltration in the tumor immune microenvironment.
Drug resistance is a primary cause of treatment failure in breast cancer. It arises from complex interactions between the tumor microenvironment, cancer cell heterogeneity, and epigenetic regulation -- dynamics that are difficult to capture with traditional approaches but well-suited to AI analysis.
A systematic test of 16 machine learning algorithms across 8 types of molecular profiles identified a four-variable classifier based on miRNA expression (hsa-miR-21-5p, hsa-miR-155-5p, hsa-miR-34a-3p, and hsa-miR-200c-3p) that predicted doxorubicin resistance with an AUC of 0.80, and model performance improved consistently as more training data was added.
A multi-omics study integrating 12 programmed cell death mechanisms across seven cohorts including TCGA (n=1,132) and METABRIC (n=1,904) constructed a 12-gene cell death index (CDI) that achieved a C-index of 0.73 to 0.81 across five independent datasets. High-CDI patients were resistant to docetaxel and oxaliplatin but sensitive to palbociclib, providing direct guidance for drug selection.
An annotation-free deep learning framework trained on routine H&E slides predicted epithelial-mesenchymal transition (EMT) phenotype and endocrine treatment response with 81.25% overall accuracy in ER-positive breast cancer, linking standard pathology images directly to resistance-relevant biological states without requiring any special molecular testing.
Despite impressive performance in research settings, several structural challenges hinder the clinical deployment of AI models in breast cancer care. Most current studies rely on single-center, retrospective datasets, which limits a model's ability to generalize to the diversity of patients, equipment, and protocols found across different healthcare systems.
The interpretability problem is particularly significant for deep learning models. Their "black box" nature -- where high predictive accuracy is achieved through patterns that cannot be easily explained -- reduces clinician trust and makes it difficult to identify when a model might be failing. Explainability tools such as feature attribution maps and saliency visualization are important but not yet standard practice.
Multimodal data integration faces practical barriers: combining genomics, imaging, and pathology data requires solving technical challenges around heterogeneous data formats, missing modalities, and the computational cost of processing large datasets from multiple sources simultaneously.
Fairness and equity concerns must also be addressed. AI models trained predominantly on certain demographic groups may underperform for underrepresented populations. Federated learning -- where models train on data distributed across multiple institutions without sharing raw patient records -- offers a promising path toward more representative and privacy-preserving model development.
The most transformative near-term opportunity lies in multimodal longitudinal integration: coupling serial mammography, ultrasound, and MRI data with digital pathology whole-slide images, bulk transcriptomics, and single-cell omics to noninvasively track tumor evolution and immune microenvironment changes over the course of treatment.
Large-scale multicenter prospective trials are essential to establish the true clinical value of AI tools. Validation endpoints should extend beyond AUC to include calibration, decision-curve net benefit, stage shift rates, positive predictive value, recall rates, and radiologist workload -- metrics that connect model scores to meaningful patient outcomes.
The development of integrated AI-assisted clinical systems that combine tumor classification, prognosis prediction, and treatment response assessment in a single user-friendly interface will be critical for adoption. Such systems must generate interpretable, actionable reports that clinicians can trust and patients can understand.
As AI technology matures and governance frameworks are established, precision oncology can shift from a vision to a routine clinical reality -- enabling earlier detection, individualized treatment selection, and ultimately improved survival and quality of life for breast cancer patients worldwide.