Diagnostic Performance and Agreement of Deep Learning- Based Digital Pathology for Liver Fibrosis Assessment: A Systematic Review and Meta-Analysis

Abstract Background Deep-learning-based digital pathology is increasingly used to stage and quantify liver fibrosis, but the evidence spans heterogeneous diseases, stains, model architectures, and outcome definitions. Methods: We systematically searched PubMed/MEDLINE, Embase, Web of Science Core Collection, Scopus, IEEE Xplore, and CNKI without year or language restrictions. Human liver pathology studies were eligible when an explicit deep-learning or neural-network method was applied to fibrosis staging, classification, agreement, detection, segmentation, or quantitative assessment. Direct staging/classification/agreement evidence was designated A-layer evidence; segmentation and continuous quantitative evidence was designated B-layer evidence. Diagnostic accuracy was pooled when compatible exact 2x2 data were available. Risk of bias and applicability were assessed using QUADAS-3. Results: Of 825 records, 28 reports representing 25 study families were included (17 A-layer and 11 B-layer reports). Four independent study families contributed exact 2x2 data for advanced fibrosis (F > = 3 or an etiologically equivalent threshold). Pooled sensitivity was 0.87 (95% CI 0.58–0.97; I2 = 64.3%) and specificity was 0.86 (95% CI 0.74–0.93; I2 = 33.3%). The pooled LR + was 6.15 (95% CI 2.96–12.78), LR- was 0.15 (95% CI 0.04–0.60), and DOR was 53.46 (95% CI 7.08-403.72). Including Pan 2026 as reported proportions produced similar sensitivity and specificity. Agreement, AUC, and B-layer outcomes were too heterogeneous for a single pooled estimate. All nine QUADAS-3 estimates had high overall risk of bias and high overall applicability concern. Conclusions: Deep-learning digital pathology shows promising performance for advanced liver fibrosis assessment, but the evidence base remains small, heterogeneous, and methodologically fragile. Multicentre external validation, locked models, standardized reporting, and prospective clinical evaluation are required before routine deployment.