Federated Learning on Heterogeneous Diffusion-Weighted Imaging Data for Acute Stroke Infarct Segmentation

In acute ischemic stroke imaging, advanced MRI techniques such as Diffusion-Weighted Imaging (DWI) and Dynamic Susceptibility Contrast Perfusion Imaging (PWI), along with threshold-based criteria, are typically used for the diagnosis and prognosis of infarcted tissue and tissue at risk. More recently, deep learning-based segmentation approaches have been proposed as an alternative to conventional thresholding-based methods; however, their utility in clinical scenarios remains limited due to the high heterogeneity of multi-centric data and the small number of labelled datasets available to train such models. Here, using a clinical dataset of 286 acute stroke patients, we first show that protocol heterogeneity is detrimental to the segmentation task with a drop of 17% in the Dice score (0.42 to 0.34) when heterogeneous data is introduced. We then introduce a simulated multi-centre scenario where each centre is limited to a small cohort of data with its own protocol, resulting in an additional overall drop of up to 16%. Finally, we show that Federated Learning, specifically the FedAvg algorithm, can be used in this heterogeneous multi-site scenario without explicit data sharing between the centres and achieve an overall Dice score comparable to the centralized model. Our results suggest that Federated Learning is a feasible alternative to data sharing for stroke segmentation, even when protocols are heterogeneous.