Federated Learning for Brain Tumor Classification from MRI: A Comparison of MLP and ConvNeXt Approaches under IID and Non-IID Data Scenarios

Federated Learning (FL) effectively addresses privacy concerns in medical imaging by enabling collaborative model training without sharing sensitive patient data. This paper compares two neural network approaches for brain tumor classification (BTC) from magnetic resonance imaging (MRI) in a federated learning setting. Both models operate on regions of interest (ROIs) extracted from brain MRI scans. The first is a lightweight multilayer perceptron (MLP) that classifies ROIs based on radiomic features extracted from them. The second is a deep learning (DL) approach based on the ConvNeXt architecture, which performs classification directly on the ROI images. Two experimental scenarios are considered: a balanced (IID) and an unbalanced (non-IID) distribution of data among federated clients. Results show that the radiomics-based MLP achieves performance comparable to the more complex ConvNeXt model, while requiring significantly lower computational resources. Moreover, federated learning consistently outperforms isolated local training, particularly under non-IID conditions, emphasizing its potential for clinical deployment.