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

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Screening mammography is a key tool for breast cancer detection.
• Accurate prediction of breast cancer molecular subtype and receptor status is crucial for treatment decisions.

Purpose of the Study:

• To systematically review and meta-analyze the performance of artificial intelligence (AI) algorithms in predicting breast cancer molecular subtype and receptor status from screening mammograms.

Main Methods:

• A PRISMA-DTA systematic review and bivariate random-effects meta-analysis was conducted.
• Searches were performed across major scientific databases (MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore).
• Twenty-five studies comparing AI mammogram predictions with histopathologic findings were included, with risk of bias and quality assessed using PROBAST and TRIPOD.

Main Results:

• AI models demonstrated moderate-to-high discrimination for predicting breast cancer subtypes, with pooled AUCs for luminal, HER2-enriched, and triple-negative subtypes ranging from 0.76 to 0.86.
• Performance for predicting specific receptor statuses varied, with AUCs for estrogen receptor (0.71), progesterone receptor (0.59), HER2 (0.64), and Ki-67 (0.60) in multi-class tasks.
• Binary receptor status prediction showed AUCs of 0.80 for HER2 and 0.71 for hormone receptor positive status.

Conclusions:

• AI applied to mammography shows promising, moderate-to-high discriminatory ability for predicting breast cancer molecular subtypes and receptor status, particularly for luminal and triple-negative disease.
• Substantial heterogeneity was observed across studies, indicating variability in AI model performance.
• Current evidence is insufficient for clinical deployment; future research should focus on larger multicenter cohorts, standardized pipelines, external validation, uncertainty quantification, and multimodal data fusion.