Deep learning-based DCE-MRI for breast lesion diagnosis: a systematic evaluation and meta-analysis

This study aimed to conduct a focused diagnostic-test-accuracy systematic review and meta-analysis of deep learning (DL)-based dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) for differentiating benign from malignant breast lesions. The specific objectives were to: [1] quantify the pooled diagnostic performance, including sensitivity, specificity, and summary receiver operating characteristic area under the curve (SROC AUC); [2] explore potential clinical, technical, and methodological sources of between-study heterogeneity; and [3] summarize the key methodological limitations of current evidence and future requirements for clinical implementation. A systematic search of the English language literature on deep learning DCE-MRI models to differentiate between benign and malignant breast lesions was conducted in PubMed, Embase, Scopus, and Web of Science databases. The search included articles from self-constructed repositories up to December 2024, and data were extracted for selected studies. Forest plots were generated using R 4.4.1 software to show the relationship between sensitivity and specificity, and subjects’ work summarized receiver operating characteristic (SROC) curves were plotted to estimate the area under the curve (AUC). From 2638 initially identified articles, 18 met the inclusion criteria. The DL-based DCE-MRI model demonstrated a pooled sensitivity of 0.90 (95% CI: 0.84–0.94) and specificity of 0.83 (95% CI: 0.78–0.87), with individual study estimates ranging from 0.64 to 1.00 and 0.66 to 0.96, respectively. The summary positive likelihood ratio was 4.69 (95% CI: 3.67–6.01), negative likelihood ratio was 0.14 (95% CI: 0.09–0.22), and diagnostic odds ratio was 40.4 (95% CI: 20.73–77.33). The overall area under the SROC curve reached 0.90 (95% CI: 0.84–0.92). DL-based DCE-MRI shows promising diagnostic accuracy for distinguishing benign from malignant breast lesions, suggesting its potential as an adjunct diagnostic tool. Due to substantial heterogeneity and limited external validation, further prospective, multi-center studies are needed before routine clinical adoption. This systematic review and meta-analysis quantitatively evaluated the diagnostic performance of deep learning-based dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) for distinguishing benign and malignant breast lesions across 18 eligible studies. The integrated model exhibited a high pooled sensitivity of 0.90, specificity of 0.83, and summary area under the SROC curve of 0.90, with country identified as the only significant source of heterogeneity. These findings suggest favorable diagnostic accuracy of DL-based DCE-MRI, but high heterogeneity and limited external validation caution against overgeneralization; prospective multi-center validation is needed prior to clinical adoption.