FedSSL: Federated Learning with Shape-Sensitive Loss for Catheter and Guidewire Segmentation

Federated learning enables multiple hospitals to train machine learning models collaboratively without sharing their sensitive local data. A significant challenge in this context is the variation in local data distributions across hospitals, where differences in patient demographics, imaging equipment, and protocols lead to inconsistencies in the data, compounded by the inherent imbalance in the number of medical images for different conditions. In this paper, we introduce a new approach, Federated Learning with Shape-Sensitive Loss (FedSSL), for catheter and guidewire segmentation. Our method enhances the federated averaging algorithm by utilizing a shape-sensitive loss function for more precise segmentation. FedSSL leverages the global model to guide local model training, mitigating the variability in local data distributions and addressing the imbalanced nature of medical images by aligning model updates to retain shape information during the training process. We conduct extensive experiments using a large-scale dataset of real and phantom X-ray images. The experimental results show that our FedSSL significantly outperforms traditional federated learning algorithms.