In-Silo Federated Learning vs. Centralized Learning for Segmenting Acute and Chronic Ischemic Brain Lesions

Abstract Purpose 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 8,421 fluid-attenuated inversion recovery images (FLAIR) from 9 hospitals for acute (Task I) and chronic (Task II) lesion segmentation. 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. We trained with datasets from 9 and 3 institutions for Task I and Task II, respectively, and externally tested them in datasets from 1 and 9 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 In Task I, the FL model employing batch clipping trained for 360 epochs achieved a DSC of 0.754±0.183, surpassing an equivalent 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 to 1.75 hours). Conclusion Comparisons between CL and FL in identical settings suggest the feasibility of FL for medical image segmentation.