Comparative Performance of Federated and Centralized Learning for Brain Tumor Classification and Segmentation

Abstract

Purpose To compare federated learning (FL) and centralized learning (CL) performance for brain tumor classification and segmentation and examine whether specific configurations confer measurable performance advantages. Materials and Methods In this systematic review and meta-analysis, Medline, EMBASE, PubMed, Web of Science, and Scopus were searched from inception to March 4, 2026. Classification accuracy was pooled using random-effects meta-analysis, with heterogeneity quantified using I 2 ; other metrics were summarized descriptively. Paired differences in classification accuracy and Dice similarity coefficient (DSC) were assessed using the Wilcoxon signed-rank test. Mixed-effects meta-regression and subgroup analyses evaluated whether specific configurations were associated with performance differences. Results Of 3,880 records identified, 50 studies were included (26 classification/detection, 26 segmentation; 2 in both). Within-study differences between FL and CL performance were -0.87 percentage points (pp) for classification accuracy (95% CI [-3.32, +1.44]; P = .41, k = 13) and -1.29 pp for segmentation DSC (95% CI [-4.23, +0.81]; Holm-corrected P = .50, k = 11). In unmatched single-arm analyses, pooled classification accuracy was 96.77% for FL (k = 18) versus 96.86% for CL (k = 9); the pooled DSC for segmentation was 84.95% (k = 24) versus 84.61% (k = 11). Heterogeneity was high (I 2 =93.3-96.5%); the 95% prediction interval for FL accuracy was 86.72-100%. Aggregation strategy, training sample size, task category, and independent and identically distributed (IID) versus non-IID partitioning were not associated with performance differences (all P ≥ .38). Conclusion There was no evidence of paired differences between FL and CL in brain tumor classification accuracy or segmentation DSC, and no specific configuration conferred a measurable performance advantage. © RSNA, 2026