This study aimed to assess how well federated learning (FL) performs compared with standard centralized learning (CL) for automatic segmentation of acute cerebral infarcts and white matter hyperintensities. We conducted a retrospective multi-center analysis involving 13,546 diffusion-weighted images (DWI) collected from 10 hospitals and 8,421 fluid-attenuated inversion recovery (FLAIR) images from 9 hospitals. These supported two separate tasks: acute lesion segmentation (Task I) and chronic lesion segmentation (Task II). Training utilized data from 9 sites for Task I and 3 sites for Task II, while external testing was performed on held-out data from 1 site (Task I) and 6 sites (Task II). During FL, the central server merged client updates every four communication rounds, applying FedYogi for Task I and FedAvg for Task II. A batch-clipping strategy was also examined in the FL framework. Segmentation accuracy was measured using the Dice similarity coefficient (DSC). Participants in the training cohorts had mean ages of 68.1 ± 12.8 years (Task I) and 67.4 ± 13.0 years (Task II), with males representing 51.5% and 60.4% of the groups, respectively. In Task I, the FL model trained for 360 epochs with batch clipping achieved a mean DSC of 0.754 ± 0.183. This outperformed a matched CL model trained for the same number of epochs (DSC: 0.691 ± 0.229; P < 0.001) and showed no significant difference from the strongest CL model trained for 940 epochs (DSC: 0.755 ± 0.207; P = 0.701). For Task II, the FL model with clipping, the FL model without clipping, and the CL model all produced statistically similar results after 48 epochs (DSCs: 0.761 ± 0.299, 0.751 ± 0.304, and 0.744 ± 0.304). Few-shot FL variants performed substantially worse. Batch clipping also halved training time for Task II (from 3.5 hours to 1.75 hours). A side-by-side evaluation of centralized and federated approaches under equivalent conditions shows that federated learning is a practical and effective approach for medical image segmentation.
