Breast MRI is the most sensitive imaging exam available for detecting breast cancer, covering the full care pathway from risk prediction and screening through treatment planning, monitoring, and follow-up. Despite its power, interpreting these scans is complex and time-consuming, and accuracy depends heavily on radiologist expertise.
Demand for breast imaging services is growing rapidly, driven by expanding clinical indications and screening guidelines for high-risk women. This pressure makes it an ideal setting for artificial intelligence (AI) to assist radiologists by handling routine triage, flagging suspicious findings, and extracting subtle patterns invisible to the human eye.
Although AI has already become mainstream in mammography through commercially available computer-aided detection (CAD) systems, its adoption in breast MRI has lagged. The primary reasons are the inherent complexity of MRI examinations and the limited availability of large, well-annotated public breast MRI datasets.
This review examines how AI techniques including radiomics, machine learning, and deep learning are being applied across the breast cancer care pathway in MRI, surveying the current evidence and discussing barriers to clinical adoption.
Radiomics extracts quantitative features from medical images, such as texture, shape, and signal intensity, that go beyond what can be seen with the naked eye. These handcrafted features can then be fed into statistical models or traditional machine learning algorithms to generate predictions about clinical outcomes.
Machine learning uses algorithms to learn patterns from data and make predictions. Traditional approaches such as support vector machines (SVMs) and random forests rely on feature engineering, where experts manually select relevant inputs. They work well on smaller datasets and produce more interpretable results than deep learning.
Deep learning, a subset of machine learning, bypasses manual feature selection by automatically identifying patterns directly from raw image data. Using convolutional neural networks (CNNs) such as ResNet, DenseNet, and U-Net, deep learning excels at complex image recognition and segmentation. More recently, Vision Transformers (ViTs) have emerged as a powerful alternative that captures global image context through self-attention mechanisms.
These three approaches can also work together. Deep learning networks can extract high-dimensional features that are combined with interpretable radiomic descriptors, with machine learning algorithms then building robust predictive models from the integrated data. This hybrid strategy leverages the strengths of each method.
A key challenge in breast MRI screening is that the vast majority of exams are negative, yet each scan produces large volumes of complex imaging data that burdens radiologists. AI-based triage systems address this by identifying likely-negative exams for expedited review and flagging suspicious ones for priority attention, improving workflow efficiency without sacrificing sensitivity.
Verburg et al. developed a deep learning triage model trained on data from over 4,500 patients in the DENSE multicenter trial. At a threshold ensuring 100% sensitivity for malignant lesions, the model achieved an AUC of 0.83 and successfully triaged nearly 40% of cancer-free scans, reducing unnecessary radiologist workload. The authors noted that accuracy decreased with higher levels of background parenchymal enhancement (BPE), a normal physiological process that can mimic cancer.
Jing et al. used a pretrained ResNet CNN fine-tuned on ultrafast MRI sequences to route patients to either an abbreviated 3-minute protocol (negative) or a full 20-minute protocol (positive). At 100% sensitivity, the model reduced radiologist workload by 6.2%. Bhowmik et al. went further, achieving 100% sensitivity across both screening and diagnostic cohorts with an 11% estimated workload reduction, while outperforming one fellowship-trained radiologist who missed a cancer obscured by marked BPE.
These studies collectively demonstrate that AI triage can match or exceed radiologist sensitivity while meaningfully reducing workload, though improving performance in the presence of dense tissue and BPE remains an active challenge.
Accurately distinguishing benign from malignant breast lesions on MRI is difficult because their imaging features often overlap, especially in dense tissue. AI models trained on multiparametric MRI data, which combines multiple sequences such as dynamic contrast-enhanced MRI, diffusion-weighted imaging, and T2-weighted imaging, consistently outperform models using any single sequence alone.
Witowski et al. used a dataset of over 21,500 breast MRI studies to train a 3D CNN that achieved an AUC of 0.92, comparable to five board-certified radiologists. Importantly, when radiologists combined their assessments with the AI model's predictions, their diagnostic accuracy and agreement improved. The model also reduced unnecessary biopsies, correctly downgrading 73.2% of benign BI-RADS 3 cases.
Luo et al. introduced MOME (Mixture-of-Modality-Experts), a large-scale transformer-based model that fuses DCE-MRI, diffusion-weighted imaging, and T2-weighted imaging using sparse modality experts and self-attention. Trained on over 5,200 patients, MOME achieved AUCs of 0.859 to 0.965 for malignancy detection, outperforming single-sequence models and demonstrating strong generalizability across external datasets.
For non-mass enhancement (NME) lesions, which are particularly challenging to classify, ResNet-50 models have achieved performance on par with senior radiologists. AI assistance was also shown to significantly boost junior radiologists' AUCs from around 0.73 to 0.86, suggesting a concrete benefit for less experienced practitioners.
Beyond classifying lesions as benign or malignant, AI is being used to characterize breast cancers in greater detail, predicting molecular subtypes such as triple-negative, HER2-positive, luminal A, and luminal B. Because these subtypes differ markedly in biological behavior and response to treatment, non-invasive subtype prediction from MRI could support faster, more personalized treatment decisions.
Multiple studies have shown that multiparametric MRI input improves subtype prediction over any single sequence. Ren et al. used an ensemble of ResNeXt50 models trained on five MRI sequences to classify all four major molecular subtypes, achieving AUCs of 0.859 to 0.912 on the combined model versus 0.611 to 0.814 for any single sequence. T1-weighted imaging proved especially effective for discriminating HER2-enriched and triple-negative cancers due to their distinct vascular profiles.
Xiang et al. introduced a dual-stream Swin Transformer network that processed two MRI sequences simultaneously, extracting both shared and modality-specific features, and incorporated receptor expression data from pathology reports processed with Google's BERT model. This multi-path structure achieved an accuracy of 85.1% and an AUC of 97.5% for molecular subtype classification, surpassing prior CNN methods.
AI has also been applied to predict histological grade, which reflects how abnormal cancer cells appear under the microscope. Sun et al. demonstrated that incorporating molecular subtype information as a secondary input into a CNN grade predictor modestly improved accuracy and AUC to 0.927, with gradient-weighted class activation maps confirming the model focused on tumor-relevant regions.
Neoadjuvant chemotherapy (NAC) is standard treatment for locally advanced breast cancer and for aggressive subtypes such as triple-negative and HER2-positive cancers. A key goal is predicting which patients will achieve a pathological complete response (pCR), meaning no residual cancer at surgery, so that therapy can be tailored and unnecessary toxicity avoided.
Radiomic models using multiparametric MRI have shown consistent ability to predict pCR. Liu et al. trained SVM models achieving an AUC of 0.79 for pCR prediction, which improved to 0.86 when combined with clinicopathologic data. For triple-negative breast cancer specifically, Hwang et al. showed that mid-treatment radiomic features outperformed baseline-only models, with AUCs of 0.78, suggesting that early imaging changes during treatment carry important predictive information.
Deep learning approaches have also proven effective. Joo et al. combined clinical and imaging data in a 3D CNN model achieving an AUC of 0.89, outperforming clinical-data-only models. A hybrid model by Comes et al. combining pretrained CNN features with clinical data achieved AUCs above 0.90. Huang et al.'s longitudinal ensemble model integrating pre- and post-NAC radiomic and deep learning features achieved AUCs of 0.959 to 0.974 across molecular subtypes.
Among the most innovative approaches, Comes et al. introduced one of the first Vision Transformer (ViT)-based models for pCR prediction, fine-tuned on just 86 patients through transfer learning. The ensemble of pre-treatment and mid-treatment ViTs achieved an AUC of 0.91 on the primary cohort and 0.81 on an independent test set, demonstrating that transformer architectures can work even with limited medical imaging data.
Traditional breast cancer risk models such as the Gail model and the Tyrer-Cuzick model rely on clinical and demographic factors that often lack precision. AI-driven approaches can instead analyze imaging-derived risk markers such as breast density, background parenchymal enhancement, and tumor microenvironment features to provide quantitative, reproducible, and more accurate risk estimates.
Portnoi et al. trained a deep learning model on screening MRI data from over 1,100 high-risk women to predict 5-year breast cancer risk, achieving an AUC of 0.64 compared to just 0.49 for the Tyrer-Cuzick model. For recurrence risk, Ha et al. used a 2D CNN to classify patients into Oncotype DX risk categories with an AUC of 0.92, matching the performance of this expensive genomic test from imaging data alone.
Fan et al. connected MRI radiomic signatures to Oncotype DX recurrence scores in estrogen receptor-positive breast cancer, showing that higher predicted scores correlated with poorer survival outcomes and greater responsiveness to NAC. This radiogenomic approach could allow recurrence risk stratification without the need for a separate genomic test.
Hirsch et al. developed a CNN pre-trained on 11,000 patients that could detect breast cancer up to one year earlier than radiologists by re-analyzing previously negative MRI scans. When reviewing the 10% highest-risk scans flagged by AI, 35 of 115 cancers were detected earlier, with AI also localizing the cancer's anatomical region in 66 cases, demonstrating the potential for truly prospective early detection.
The most significant barrier to clinical AI adoption is the black box problem: deep learning models often cannot explain why they made a particular prediction. This opacity erodes clinician trust, especially in high-stakes scenarios like cancer diagnosis. Efforts to improve explainability through attention visualization, feature attribution maps, and post-hoc explanation techniques are underway but not yet standard.
Accountability and legal liability present another major hurdle. If an AI system misses a cancer or incorrectly recommends a biopsy, it is unclear who bears responsibility. Clear regulatory frameworks defining accountability for AI developers, healthcare institutions, and radiologists are needed before broad clinical deployment can proceed responsibly.
Automation bias is a subtle but important risk: studies show that radiologists at all experience levels can over-rely on AI recommendations, sometimes accepting incorrect AI outputs without sufficient critical evaluation. Education programs that frame AI as a decision-support tool supplementing but not replacing human judgment are essential to mitigate this effect.
Technical challenges include integrating AI into existing Picture Archiving and Communication Systems (PACS), ensuring data privacy under regulations such as HIPAA and GDPR, and achieving model generalizability across different hospitals, scanner types, and patient populations. Federated learning, which trains models across distributed datasets without centralizing sensitive data, offers a promising path to building larger training sets while maintaining privacy.
The field of AI in breast MRI is rapidly maturing, with demonstrated value across screening triage, lesion classification, molecular subtyping, treatment response prediction, and risk assessment. As breast MRI becomes more widely indicated, AI-driven tools will be essential for managing radiologist workload while maintaining or improving diagnostic quality.
Future advances will likely include standardized validation frameworks and regulatory pathways to ensure AI tools are rigorously tested before clinical deployment. Dual-review systems, in which AI outputs are independently verified by radiologists, can help prevent automation bias while still delivering efficiency gains.
Continual learning architectures will allow AI models to update as imaging protocols and patient populations evolve without losing previously learned knowledge. AI will increasingly move from general breast cancer predictions toward subtype-specific risk assessments, enabling more targeted screening and prevention strategies for patients at highest risk of aggressive cancers such as triple-negative breast cancer.
The convergence of AI with breast MRI holds transformative promise for personalized cancer care. By addressing current technical, ethical, and regulatory barriers through collaborative effort among radiologists, AI developers, regulators, and patients, AI-enhanced breast MRI has the potential to improve early detection, reduce overtreatment, and ultimately improve outcomes for patients with breast cancer worldwide.