The clinical rationale for AI augmentation is stark. Perceptual errors account for an estimated 60 to 80 percent of missed diagnoses in retrospective audits, and fatigue during high-volume reporting can reduce radiologist performance. Inter-reader variability among experienced radiologists, with Cohen’s kappa around 0.4 to 0.6, affects recall and biopsy thresholds. Meanwhile, mammographic screening sensitivity falls to 48 percent in dense breasts, and false-positive rates of 10 to 12 percent per round generate unnecessary procedures and patient anxiety. Five-year relative survival approaches 99 percent for localized disease but drops to 28 percent for distant disease, making early detection central to improving outcomes.

The methodological landscape covered by the review spans classical machine learning and radiomics, convolutional neural networks, YOLO-family detectors, vision transformers and hybrid CNN-Transformer architectures, U-Net variants, Mask R-CNN, SAM, and MedSAM across mammography, tomosynthesis, MRI, ultrasound, contrast-enhanced mammography, and emerging photoacoustic imaging. Convolutional neural networks have dominated breast imaging AI since the ImageNet deep-learning breakthrough, exploiting spatial structure through hierarchical local feature extraction. Single-stage YOLO detection frameworks formulate lesion localization as a direct regression problem, enabling fast inference suitable for real-time detection workflows. Vision transformers apply global self-attention to sequences of image patches, motivating hybrid CNN-Transformer approaches for medical image analysis.

Representative CNN studies show impressive retrospective benchmarks. A deep-learning mammography system evaluated on independent UK and US screening cohorts achieved an AUC of 0.871 in a reader study compared with a mean reader AUC of 0.750, with absolute reductions in false positives of 1.2 percent in the UK and 5.7 percent in the US, and absolute reductions in false negatives of 2.7 and 9.4 percent respectively. A simulated UK double-reading workflow maintained non-inferior performance while reducing second-reader workload by 88 percent. The open-source Mirai model, developed at Massachusetts General Hospital and evaluated on held-out cohorts from Sweden and Taiwan, obtained C-indices of 0.76 to 0.81 for one- to five-year breast cancer risk prediction, outperforming established clinical risk models where direct comparison was available.

YOLO-family models have shown promising retrospective evidence for fast detection and localization. One early YOLO-based CAD system using 600 original DDSM mammograms and 2,400 augmented images achieved 99.7 percent mass-detection accuracy and 97.0 percent benign/malignant classification accuracy under five-fold cross-validation, although the review cautions that such very high accuracies on single benchmarks or heavily augmented data carry substantial overfitting risk and do not equate to clinical sensitivity or specificity in prospective screening workflows. A YOLOv5-CAD framework using transfer learning across CBIS-DDSM, VinDr-Mammo, INbreast, MIAS, and a private cohort achieved mAP 0.843 with 14.8 millisecond inference time when transferring from CBIS-DDSM to INbreast, with external MIAS and private cohort mAP values of 0.756 and 0.816 respectively.

Segmentation architectures have evolved from standard U-Net through attention mechanisms, transformer integration, and foundation model-based promptable segmentation. The original U-Net, introduced in 2015 with its symmetric encoder-decoder structure and skip connections, achieved Dice of 0.92 on histological cell segmentation and has been forked over 10,000 times on GitHub. Attention U-Net applied to the BUSI ultrasound dataset achieved Dice of 0.87 and Hausdorff distance of 5.8 millimeters for breast lesion boundary delineation. The nnU-Net framework applied to 100 three-dimensional fat-suppressed breast DCE-MRI scans achieved Dice of 0.968 for whole breast and 0.877 for fibroglandular tissue. MedSAM, fine-tuned on 1.5 million medical image-mask pairs spanning 10 imaging modalities, achieved Dice of 0.88 on the breast imaging subset. Mammo-SAM, a breast-specific adaptation of the Segment Anything Model, improved breast mass segmentation over unadapted SAM on CBIS-DDSM and INbreast, demonstrating that foundation segmentation models require domain adaptation before clinical use in mammography.

Explainable AI has become essential because diagnostic models must support radiologist trust, error analysis, and regulatory transparency under the EU AI Act and FDA guidance. The review highlights Grad-CAM and Grad-CAM++ saliency heatmaps, bounding-box and lesion-localization outputs, segmentation masks, attention visualization, and case-based interpretability as the explainability approaches most relevant to breast imaging. Grad-CAM generates class-discriminative heatmaps by weighting convolutional feature maps using gradient information, helping radiologists assess whether model attention overlaps with suspicious masses, calcifications, or enhancement regions. Case-based interpretable models support predictions by comparison with visually similar mammographic mass cases, a clinically intuitive approach because radiologists often reason by comparison with prior cases. The review emphasizes that localization maps and segmentation masks should not be interpreted as full mechanistic explanations, and attention weights should be interpreted cautiously because they do not always provide faithful explanations of model causality.

Multimodal fusion, radiogenomics, and federated learning represent emerging directions. A machine-learning radiogenomic analysis using 922 patients and 529 DCE-MRI features demonstrated moderate associations between MRI-derived quantitative features and molecular markers including ER, PR, HER2, Ki-67, and molecular subtype. Federated learning addresses the privacy challenge of building generalizable models across institutions without centralizing sensitive patient data, sharing model updates rather than raw imaging data. Public datasets including CBIS-DDSM, INbreast, VinDr-Mammo, BUSI, Duke Breast Cancer MRI, and EMBED, with approximately 3.4 million mammograms and linked demographic metadata, enable reproducible benchmarking and fairness auditing, while open-source systems such as Mirai, nnU-Net, and MedSAM allow local validation.

Recent prospective and implementation evidence strengthens the clinical-translation component. The final MASAI randomized trial, including more than 105,000 women, showed that AI-supported screening achieved a non-inferior interval cancer rate, higher sensitivity, unchanged specificity, and 12 percent interval-cancer reduction versus standard double reading. The AI-STREAM prospective multicenter cohort in South Korea, evaluating 24,543 women, reported a 13.8 percent increase in cancer detection without a significant recall-rate increase. The AITIC paired noninferiority trial showed that partially autonomous AI triage reduced radiologist workload by 63.6 percent while increasing cancer detection by 15.2 percent, although recall rates increased in the overall cohort and did not meet the prespecified noninferiority criterion. Differences in recall burden probably reflect threshold setting, whether AI was used as triage or decision support, and local workflow design.

Despite these advances, persistent challenges include data heterogeneity, class imbalance, domain shift, bias, incomplete calibration, and limited prospective validation. The review found that only 38 of the 284 evidence-map entries were verified from full text, 48 were based on abstracts, and 198 were based on DOI or source pages, and the authors emphasize that the supplementary data sheet should be interpreted as a provenance-tagged evidence map rather than a formal study-level risk-of-bias assessment. The most frequent methodological concerns were retrospective single-institution development, internal-only train-test splits with possible patient or lesion-level leakage, limited independent external validation, scanner and vendor-related domain shift, and incomplete reporting of calibration, subgroup performance, and failure cases. The authors conclude that broader clinical translation requires robust external and prospective validation, improved calibration, assessment of generalizability and bias, and integration into clinical workflows, with future research prioritizing prospective multicenter trials, lesion-level evaluation, subgroup fairness analysis, uncertainty quantification, and calibration across diverse populations and imaging systems.

Subject of Research: Artificial intelligence in breast cancer imaging

Article Title: Artificial intelligence in breast cancer imaging: a systematic review of detection, segmentation, explainability, and clinical translation

Article References: Khan, I. I., Shah, S. A. H., Tsipourakis, A., Malavolta, M., Aleesha, A., Shah, S. B. H., Hajiani, M., Buccoliero, A., Panagiotopoulos, K., Shah, S. T. H., & Deriu, M. A. (2026). Artificial intelligence in breast cancer imaging: a systematic review of detection, segmentation, explainability, and clinical translation. Frontiers in Imaging, 5, Article 1910013. https://doi.org/10.3389/fimag.2026.1910013

Image Credits: AI Generated

DOI: 10.3389/fimag.2026.1910013

Keywords: artificial intelligence, breast cancer, mammography, deep learning, medical imaging, image segmentation, explainable AI, clinical translation, systematic review, radiomics, federated learning, vision transformers