Аннотация
Early diagnosis and therapy planning in brain tumor treatment depend heavily on medical image segmentation. However, Traditional DNNs require data from medical practices to be collected in a central location, raising significant privacy concerns about patient data leakage, security risks, and institutional data-sharing limits. This study proposes a robust, multi-institutional medical image segmentation system that uses federated deep learning while maintaining the privacy of clinical data. The proposed system combines a U-Net-based convolutional neural network with a federated learning architecture. The goal is to obtain multi-institution segmentation models without directly sharing healthcare institutions' original MRI data. To enhance data protection, differential privacy measures, such as gradient clipping and the injection of Gaussian noise, were applied during model optimization and the exchange of model parameters. The brain tumor MRI dataset from BraTS 2020 was used for experiments and distributed across four simulated healthcare clients to mimic diverse institutional environments. The experimental implementation was carried out using Python and PyTorch on Google Colaboratory. The segmentation results were evaluated with Dice Similarity Coefficient (DSC) and Intersection over Union (IoU). The Dice and IoU values were 0.3628 and 0.2464 for the centralized U-Net, and 0.3295 and 0.2176 for the federated U-Net, respectively, indicating that the federated U-Net achieved the highest segmentation performance while maintaining effective collaborative learning and privacy preservation. While the DP-Federated model offered stronger privacy protection, segmentation accuracy was compromised, with lower Dice and IoU scores of 0.0408 and 0.0211, respectively, due to noise perturbation. The results suggest a federated learning model with suitable segmentation capabilities and a significant reduction in exposure to sensitive medical data.