A Privacy-Protected Federated Learning with Cross-silo Brain Tumour Dataset for Glioma Detection

Brain tumours are abnormal growths of cells within the brain or the central spinal canal. The occurrence of this disease in a critical location can cause significant neurological complications. Accurate and early brain tumour detection is required for effective treatment planning and improving patient health. Brain tumours vary widely in type and severity, making it complex to detect and characterize through imaging techniques. The deep learning models learn to identify patterns associated with brain tumours from large volumes of imaging data, leading to high levels of accuracy in diagnosis. However, traditional deep learning models require centralized collection and processing of data, which can raise significant privacy concerns. Additionally, assembling a large amount of data at a centralized location can also be challenging and costly. The proposed work implements a federated learning-based InceptionV3 model for the detection of brain tumours. The objective is to leverage decentralized data sources while preserving privacy and enhancing the model’s ability to generalize accurate outcomes across diverse datasets. The proposed federated learning framework has been implemented with the InceptionV3 technique in independent and identically distributed (IID) and non-IID distributions for analyzing the performance across multiple rounds. The results demonstrate a remarkable performance in brain tumour detection by achieving an accuracy of 98.01% and 97.44% with IID and non-IID distributions, respectively. These findings validate the potential of FL-based approaches in medical imaging, particularly using the InceptionV3 model, to achieve high prediction performance while adhering to data privacy constraints.