Abstract
Brain tumors are an extremely deadly condition and the growth of abnormal cells that have formed inside the brain causes the illness. According to studies, Magnetic Resonance Imaging (MRI) is a fundamental imaging method that is frequently used in medical diagnostics to identify, treat, and routinely check for brain cancers. These images include extremely private and delicate details regarding the brain health of the individuals and it must be treated with much care to ensure anonymity of patients. However, traditional brain tumor segmentation techniques usually rely on centralized data storage and analysis, which might result in privacy issues and violations. Federated learning offers a solution by enabling the cooperative development of brain tumor segmentation models without necessitating the transfer of raw patient data to a centralized location. All the data are held securely within their institution. A Reinforcement Learning-based Federated Averaging (RL-FedAvg) model is proposed that fuses the Federated Averaging (FedAvg) model with Reinforcement Learning (RL). To optimize the global model for image segmentation jobs as well as to govern the consumption of client resources, the model dynamically updates client hyperparameters upon real-time performance feedback. A Double Attention-based Multiscale Dense-U-Net model, known as mixed-fed-UNet, is proposed in the work that uses the RL-FedAvg algorithm. The proposed technique achieves 98.24% accuracy and 93.28% dice coefficient on BraTs 2020 dataset. While comparing the developed model with the other existing methods, the proposed methodology shows better performance.