Diagnostic Accuracy of Artificial Intelligence-Assisted Mammography Interpretation Vs. Radiologist Alone: A Systematic Review and Meta-Analysis

Abstract

Background: Breast cancer has the highest mortality rate among women worldwide, and early detection through mammography is critical for reducing breast cancer mortality. Artificial intelligence (Al) has become a promising tool to support radiologists in mammography interpretation. However, the diagnostic accuracy of AI-assisted interpretation needs to be evaluated and compared with current methods. Objectives: This systematic review and meta-analysis aimed to evaluate the diagnostic accuracy of AI-assisted mammography interpretation versus radiologist-alone interpretation in detecting breast cancer. Methods: A comprehensive literature search identified studies that compared AI-assisted mammography interpretation with radiologist-alone assessment. Twenty-five eligible studies were included, encompassing retrospective cohorts, prospective trials, and multi-reader studies across diverse healthcare settings and AI algorithms. Data were extracted for sensitivity, specificity, likelihood ratios, and diagnostic odds ratios. Pooled estimates were calculated using a bivariate I² statistic, and heterogeneity was assessed with variance measures, including median odds ratios. The risk of bias was evaluated using ROBINS-I and QUADAS-2 tools. Results: The pooled sensitivity of AI-assisted interpretation was 0.82 (95% CI: 0.77–0.85), and specificity was 0.90 (95% CI: 0.85–0.93). The area under the SROC curve (AUC) was 0.91 (95% CI: 0.88–0.93), demonstrating high overall diagnostic performance. The pooled Positive likelihood ratio (PLR) was 7.9 (95% CI: 5.2–11.8), the negative likelihood ratio (NLR) was 0.21 (95% CI: 0.16–0.26), and the diagnostic odds ratios (DOR) were 38 (95% CI: 22–66). These findings suggest that AI significantly enhances the sensitivity of breast cancer detection while maintaining high specificity. However, substantial heterogeneity was observed across studies, reflecting differences in populations, algorithms, and thresholds. Conclusion: AI-assisted mammography interpretation comparable to radiologists alone demonstrates high diagnostic accuracy, which may improve early detection rates and reduce false positives. It has the potential to augment radiologist performance in breast cancer screening and integrate into larger-scale breast cancer screening programs. While these results support integration of AI into clinical workflows, further prospective, real-world validation with standardized thresholds is required before widespread adoption.