Performance evaluation of deep learning algorithms in MRI breast lesion segmentation and detection

Purpose

This study systematically evaluates the efficacy of deep learning (DL) algorithms for segmenting and detecting breast lesions in magnetic resonance imaging (MRI), focusing on segmentation accuracy and clinical applicability.

Methods

Following PRISMA-DTA guidelines, we searched PubMed, Embase, Scopus, and Web of Science, identifying 19 eligible studies. Inclusion criteria included MRI studies using DL for breast lesion segmentation and detection, with comprehensive data on segmentation efficacy. Study quality was assessed using QUADAS-AI. Meta-analysis was performed using random-effects modeling, with segmentation accuracy quantified by the Dice similarity coefficient (DSC) and lesion detection efficacy by sensitivity. Heterogeneity was explored through meta-regression and subgroup analysis.

Results

The 19 studies evaluated DL algorithms like U-Net, nnU-Net, and CNN. DSC for segmentation ranged from 0.61 to 0.97, with a pooled DSC of 0.82 (95 % CI: 0.76–0.88). Pooled sensitivity across six studies was 0.86 (95 % CI: 0.75–0.98). Subgroup analyses showed higher accuracy in multicenter studies (0.86 vs. 0.80), studies with external validation (0.89 vs. 0.79), and 3.0 T MRI devices (0.88 vs. 0.83). Intensity normalization also improved accuracy (0.87 vs. 0.79). nnU-Net achieved the highest DSC (0.97). Significant heterogeneity (I² = 99.6 %) and publication bias (p = 0.018) were observed.

Conclusion

DL algorithms show high accuracy in breast lesion segmentation and detection, particularly in multicenter studies and those with external validation. Future research should optimize algorithms to reduce heterogeneity and validate clinical applicability.