In-silo federated learning vs. centralized learning for segmenting acute and chronic ischemic brain lesions

Objectives: To investigate the efficacy of federated learning (FL) compared to industry-level centralized learning (CL) for segmenting acute infarct and white matter hyperintensity. Materials and methods: This retrospective study included 13,546 diffusion-weighted images (DWI) from 10 hospitals and 8421 fluid-attenuated inversion recovery (FLAIR) images from 9 hospitals for acute (Task I) and chronic (Task II) lesion segmentation. We trained with datasets originated from 9 and 3 institutions for Task I and Task II, respectively, and externally tested them in datasets originated from 1 and 6 institutions each. For FL, the central server aggregated training results every four rounds with FedYogi (Task I) and FedAvg (Task II). A batch clipping strategy was tested for the FL models. Performances were evaluated with the Dice similarity coefficient (DSC). Results: The mean ages (SD) for the training datasets were 68.1 (12.8) for Task I and 67.4 (13.0) for Task II. The frequency of male participants was 51.5 % and 60.4 %, respectively. In Task I, the FL model employing batch clipping trained for 360 epochs achieved a DSC of 0.754 ± 0.183, surpassing an equivalently trained CL model (DSC 0.691 ± 0.229; p < 0.001) and comparable to the best-performing CL model at 940 epochs (DSC 0.755 ± 0.207; p = 0.701). In Task II, no significant differences were observed amongst FL model with clipping, without clipping, and CL model after 48 epochs (DSCs of 0.761 ± 0.299, 0.751 ± 0.304, 0.744 ± 0.304). Few-shot FL showed significantly lower performance. Task II reduced training times with batch clipping (3.5–1.75 h). Conclusions: Comparisons between CL and FL in identical settings suggest the feasibility of FL for medical image segmentation.