Deep Learning Algorithms Versus Radiologists in Digital Breast Tomosynthesis for Breast Cancer Detection: Systematic Review and Meta-Analysis.

BACKGROUND

Deep learning (DL) algorithms for digital breast tomosynthesis (DBT) have proliferated, demonstrating emerging potential in enhancing lesion detection and classification.

OBJECTIVE

This study aimed to compare the diagnostic performance of DL algorithms for DBT with that of radiologists of varying experience and assess the clinical impact of DL assistance.

METHODS

A systematic search of PubMed, Embase, Web of Science, and the Cochrane Library was conducted up to November 8, 2025. Included studies compared the performance of stand-alone DL algorithms for DBT, radiologist interpretation alone, and DL-assisted diagnosis. Study quality was assessed using the Prediction Model Risk of Bias Assessment Tool+Artificial Intelligence (PROBAST+AI). Performance metrics were pooled using bivariate random effects and generalized linear mixed models.

RESULTS

A total of 13 studies with 38,565 patients were included in the final analysis. Stand-alone DL algorithms achieved a pooled sensitivity of 0.88 (95% CI 0.80-0.93), specificity of 0.74 (95% CI 0.59-0.85), and area under the receiver operating characteristic curve (AUC) of 0.89 (95% CI 0.86-0.92). While DL performance showed no statistically significant difference compared to all radiologists (AUC=0.89 vs 0.88; P=.64) or senior radiologists (AUC=0.89 vs 0.90; P=.48), DL demonstrated significantly superior sensitivity compared to junior radiologists (0.88 vs 0.76; P=.03). Notably, DL assistance did not statistically improve diagnostic metrics for radiologists across any experience level. Meta-regression identified validation methods as a significant source of heterogeneity.

CONCLUSIONS

DL algorithms for DBT exhibited strong diagnostic proficiency and showed higher sensitivity than junior radiologists, suggesting their potential utility as adjunctive tools to help reduce oversight in less experienced settings. However, given that DL assistance did not significantly elevate overall human performance, current models act primarily as supplementary aids rather than definitive clinical tools. Future prospective multimodal studies are warranted to validate these findings and optimize clinical integration.