Non-Invasive Breast Cancer Receptor Typing from Mammograms Using Artificial Intelligence: A Systematic Review and Meta-Analysis

Artificial intelligence (AI) applied to screening mammography may non-invasively predict breast cancer molecular subtype and receptor status. We conducted a PRISMA-DTA systematic review and bivariate random-effects meta-analysis (PROSPERO CRD420251032810) on this subject. Methods: We conducted a thorough search in MEDLINE, Embase, Scopus, Web of Science, and IEEE Xplore up to May 2025. Eligible studies compared mammogram AI predictions with histopathologic findings. Risk of bias was assessed with the PROBAST tool, and quality assessment was done using transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD). Twenty-five studies met the inclusion criteria. On internal test sets, pooled AUC/sensitivity/specificity were 0.86/84%/80% for luminal subtype, 0.80/70%/78% for HER2-enriched tumors, and 0.76/75%/83% for triple-negative breast cancer. Multi-class receptor-status tasks yielded AUCs: estrogen receptor 0.71, progesterone receptor 0.59, HER2 0.64, and Ki-67 0.60. Binary receptor-status tasks provided AUCs: HER2 0.80 and hormone receptor positive 0.71. Heterogeneity was substantial (I2 often > 75%). AI from mammograms shows moderate-to-high discrimination, strongest for luminal and triple-negative disease, but evidence is insufficient for clinical deployment. Priorities include larger multicenter cohorts, standardized pipelines, preregistered external validation, uncertainty quantification, and multimodal fusion.