An interpretable approach for schizophrenia classification using fMRI and sMRI features

Schizophrenia is a neurodivergent disorder that can be studied using neuroimaging-based machine learning models for early diagnosis and classification. Despite advances in neuroimaging, a gap remains in visualising multimodal magnetic resonance imaging (MRI) data, compounded by challenges in interpretability and complex feature extraction. In this study, we propose a novel framework that integrates structural MRI (sMRI) and functional MRI (fMRI) data to improve schizophrenia classification accuracy while visualizing disorder-specific abnormalities. We employed the MLSP 2014 Schizophrenia Classification Challenge dataset comprising 86 subjects (40 schizophrenia patients and 46 healthy controls) and extracted 410 neuroimaging features, 378 FNC features from fMRI, and 32 SBM features from sMRI, using independent component analysis (ICA). To enhance clinical relevance, we further validated our approach on the publicly available COBRE (Center for Biomedical Research Excellence) dataset, which provides high-resolution anatomical MRI and resting-state fMRI scans for 147 participants (72 schizophrenia patients and 75 healthy controls) along with phenotypic details such as age, gender, handedness, and diagnostic information. The proposed multi-scale recurrent neural network (MsRNN) achieved 83.33% accuracy on the MLSP dataset and 89.8% on the COBRE dataset. To improve interpretability, layer-wise relevance propagation and gradient-weighted class activation mapping generated clinically meaningful visualizations of discriminative brain regions. These results demonstrate that the proposed multimodal, XAI-integrated framework outperforms conventional models while offering transparent, clinically useful explanations to aid diagnostic decision-making.