A federated stroke segmentation to impact limited data institutions.

Abstract

Stroke, predominantly caused by blood vessel occlusion, is the second leading cause of death worldwide. DWI sequences facilitate characterization of brain-affected tissue, enabling lesion volume estimation, guiding treatment protocols, and aiding in prognosis approximation. However, radiological interpretations rely on neuroradiologist expertise, introducing subjectivity. Currently, computational solutions have allowed to support lesion characterization, but such efforts are dedicated to learn patterns from only one institution, lacking the variability to generalize geometrical lesion shape models. Moreover, some institutions lack training samples in annotated batches, which makes it difficult to achieve personalized solutions. This work introduces the first federated approach to stroke segmentation, leveraging data across institutions to impact institutions without data requirements. Models were trained on diverse institutional data and combined to obtain a robust solution for those without annotated datasets. Also, from such federated scheme was possible to measure the generalization capability of state-of-the-art architectures, evidencing new challenges in stroke care support.Clinical relevance- The validation of federated collaborative solutions to support stroke segmentations to transfer in clinical scenarios.