This review article surveys the state of deep learning across the full spectrum of breast imaging: mammography, digital breast tomosynthesis (DBT), ultrasound, and MRI. It covers both what has already been achieved clinically and what remains under development.
Deep learning has transformed radiology by enabling computers to extract meaningful patterns from medical images. In breast imaging, where early detection is directly linked to survival, AI tools have attracted enormous research interest and are beginning to enter clinical practice.
The review was published in BJR Open in 2022 and provides a structured overview aimed at radiologists and clinicians who want to understand the current landscape and realistic potential of AI in breast imaging.
Breast cancer screening generates enormous volumes of imaging data - millions of mammograms are taken every year. This scale makes breast imaging one of the richest domains for training AI systems. Large, well-curated datasets from screening programs have enabled rapid development of high-performing models.
Despite high-volume screening, breast cancer remains difficult to detect reliably. Sensitivity of mammography interpretation varies between radiologists, and up to 20% of cancers are missed on initial screening reads. AI tools offer the potential to reduce this variability and catch cancers that human readers overlook.
The field has benefited from the availability of public benchmarked datasets and AI competitions (such as DREAM and RSNA challenges), which have accelerated model development and provided standardized performance comparisons.
In 2D mammography, AI systems have been developed for cancer detection, mass characterization, calcification detection, and density assessment. Multiple commercial AI products have received regulatory clearance and shown in reader studies that they can match or supplement radiologist performance.
For calcifications specifically - clusters of tiny calcium deposits that can indicate early breast cancer - AI has demonstrated particular promise. Calcifications can be subtle and easy to miss, making them a strong use case for AI-assisted flagging.
In DBT (3D mammography), AI addresses a key limitation: the volume of images per patient is 3-4 times larger than 2D mammography, increasing reading time significantly. AI triage and detection tools reduce this burden by pre-marking suspicious slices and masses for radiologist review.
Breast ultrasound is widely used as a complement to mammography, particularly for women with dense breast tissue. AI systems have been developed to classify ultrasound masses as benign or malignant, with some studies showing AI-assisted interpretation reduces unnecessary biopsies.
Breast MRI is the most sensitive breast imaging modality and is used for high-risk screening and treatment planning. AI applications include lesion detection, characterization, and predicting response to neoadjuvant chemotherapy (treatment given before surgery).
A particular focus of MRI AI research is predicting pathological complete response (pCR) - whether the tumor disappears entirely after pre-surgery chemotherapy. Accurate pCR prediction could allow surgeons to tailor the extent of surgery, potentially avoiding unnecessary removal of breast tissue.
Beyond detecting existing cancers, AI is being applied to risk assessment - predicting which women are most likely to develop breast cancer in the future. Models trained on mammographic features have shown ability to predict cancer risk 1-5 years before a tumor is detectable, potentially enabling more personalized screening intervals.
AI also has potential to optimize radiology workflows. By triaging studies (flagging high-suspicion cases for priority reading), pre-populating radiology reports, and filtering out clearly normal studies, AI could reduce radiologist workload and improve reading efficiency.
These workflow applications may be where AI creates the most immediate clinical impact - not by replacing radiologists, but by allowing them to focus their attention where it matters most.
Despite impressive research results, widespread clinical adoption of AI in breast imaging faces significant challenges. Most published studies are retrospective, use single-institution datasets, and measure AI performance in isolation rather than in the context of how radiologists actually use it.
Generalization is a recurring concern: models trained on data from one scanner model or patient population often perform worse on data from different sites. Prospective, multi-site validation studies are needed before most tools should be trusted in clinical practice.
Regulatory frameworks for AI medical devices are still evolving, and questions around liability, reimbursement, and integration with existing radiology information systems present practical barriers that technical performance alone cannot address.
Deep learning in breast imaging has transitioned from a research curiosity to a clinically active field. Multiple commercial AI tools are now FDA-cleared and deployed in real screening programs, with demonstrated impact on radiologist performance and workflow.
The next frontier involves moving beyond detection toward prediction - not just where is the cancer, but how aggressive is it, how will it respond to treatment, and who is most at risk. These are harder problems that will require integration with clinical data beyond imaging.
The review concludes with an optimistic but cautious outlook: AI has genuine and growing utility in breast imaging, but rigorous prospective evaluation in diverse real-world settings remains essential before its full potential can be realized.