Enhancing Parallelism in Cross-silo Federated Learning for Brain Disease Classification

The brain is essential for regulating various bodily processes, and preserving its well-being is crucial for a good quality of life. Disorders like Alzheimer’s disease and brain tumors significantly impact individuals’ lives. Traditional centralized learning methods, such as machine learning and deep learning, have demonstrated improved performance in detecting these diseases. However, a major concern in the present day is the inability of centralized learning to protect the privacy of subjects’ data. To address it, this paper focuses on privacy-preserving cross-silo federated learning for detecting Alzheimer’s disease and brain tumors. The LeNet5 model, with the Adam and SGD optimizers, analyzes T1-weighted magnetic resonance imaging data and achieves an average of 95% accuracy, precision, recall, and F1 score for Alzheimer’s disease classification with 90% and 100% client participation. For brain tumor classification, a mean accuracy of 94.66%, and precision, recall, and F1 score of 95% are achieved using the LeNet5 with SGD optimizer. Our findings also manifest that it is very challenging to determine the percentage of client participation required to achieve a moderate or better performance in cross-silo federated learning since the result fluctuates but it is clear that about a minimum of 20% of client participation is required in case of balanced or mostly balanced dataset.