Breast MRI (magnetic resonance imaging) is the most sensitive breast imaging technique available, capable of detecting cancers missed by mammography and providing detailed information about tumor extent, biology, and treatment response. However, it is also the most complex to interpret - generating large volumes of data with dynamic contrast enhancement patterns that challenge even experienced radiologists.
This review, published in the Journal of Magnetic Resonance Imaging in 2024, maps out current and emerging AI applications for breast MRI, from automated detection through treatment response monitoring and future integration with genomics.
The review is designed for breast radiologists and oncologists seeking to understand which AI tools are ready for clinical use today and which represent the next wave of development.
A standard breast MRI protocol generates hundreds of images per patient across multiple sequences (T1, T2, dynamic contrast-enhanced). Dynamic contrast-enhanced MRI captures how quickly contrast agent enters and leaves tissue, which reflects tumor vascularity - a hallmark of malignancy. Analyzing all these sequences requires significant time and expertise.
Breast MRI is recommended for high-risk women (those with BRCA gene mutations or strong family history), for staging newly diagnosed cancers, and increasingly for monitoring response to neoadjuvant chemotherapy. In all these settings, the volume of data and interpretive complexity create opportunities for AI assistance.
Inter-reader variability in breast MRI interpretation is well-documented: radiologists frequently disagree on lesion characterization and size measurements. AI provides a consistent, quantitative measurement system that does not vary between readers or across reading sessions.
AI systems for breast MRI lesion detection work by automatically identifying suspicious enhancing lesions in the 3D volume, reducing the risk of missed lesions and speeding up the initial review. In studies of high-risk screening MRI, AI detects cancers with sensitivity comparable to experienced breast radiologists.
Beyond detection, AI can classify lesions as likely benign or malignant. Features such as kinetic enhancement curves (how contrast washes in and out), morphology, and texture are quantified by AI and mapped to probability scores. These scores can reduce unnecessary biopsies of benign lesions.
AI is also being applied to background parenchymal enhancement (BPE) - the normal enhancement of breast tissue that can obscure tumors. Automated BPE quantification gives a consistent measurement that correlates with breast cancer risk and affects how screening MRI is interpreted.
Neoadjuvant chemotherapy (chemotherapy given before surgery) is increasingly standard for locally advanced breast cancer. Whether the tumor completely disappears (pathological complete response, pCR) is one of the strongest predictors of long-term survival. MRI is used to monitor this response, but accurate measurement of residual tumor extent is challenging.
AI tools for tumor volume measurement from MRI can track changes in tumor size between treatment cycles with greater precision and reproducibility than manual measurement. Some studies suggest AI-measured volume change after the first treatment cycle can predict pCR before treatment is completed.
Early, accurate response prediction has major clinical implications: it could allow clinicians to switch ineffective chemotherapy regimens mid-course rather than completing a full cycle that is not working, potentially improving outcomes and reducing toxicity.
MRI-based AI risk models analyze breast tissue characteristics visible on MRI - density, fibroglandular tissue patterns, BPE - to estimate a woman's breast cancer risk over the next several years. These models complement genomic tests like BRCA testing.
Identifying women who are at high risk but who would not be captured by genetic testing alone is a major gap in current screening guidelines. AI risk models derived from imaging could identify additional women who would benefit from annual MRI screening.
The combination of AI risk scores from imaging with clinical factors and genetic information represents a multi-modal risk model that may prove more accurate than any single risk assessment approach.
A significant challenge for breast MRI AI is that image quality varies substantially between scanner manufacturers, field strengths, and acquisition protocols. AI models trained on data from one institution may perform poorly on data from another if the image characteristics differ.
AI is being applied to the acquisition process itself - techniques like deep learning-based image reconstruction can improve image quality, reduce scan time, or enable lower contrast agent doses. AI denoising of undersampled MRI data (acquired faster) to produce quality equivalent to fully-sampled data is an active research area.
Standardization efforts - developing AI tools that are robust to scanner variability - are essential for enabling multi-center deployment of breast MRI AI without requiring retraining at each institution.
The review envisions a future breast MRI workflow where AI performs automated lesion segmentation, volumetric measurement, kinetics analysis, and risk scoring simultaneously - delivering a structured quantitative report to the radiologist who then adds clinical interpretation.
Integration with pathology (genomics, biomarkers) and clinical data (treatment history, family history) within AI models is the next step. These multimodal models will likely outperform imaging-only AI by incorporating the full context of each patient's case.
The review concludes that breast MRI AI is already delivering clinical value in select applications and that the pace of progress suggests comprehensive AI-assisted breast MRI workflows will be standard of care within the next decade.