Anatomy-Aware Deep Learning Approach for Schizophrenia Diagnosis Using Structural MRI

Schizophrenia is a persistent mental condition which is characterized by severe structural and functional impairment in particular systems in the brain, which usually results in cognitive, emotional and behavioural disorders. Extensive use of traditional diagnosis is based on the subjective assessment of the symptoms, and it can lead to the late or inaccurate determination of the disorder. To cope with this issue, the current paper will suggest an effective hybrid model of deep learning that incorporates both anatomical segmentation and classification of identifying the presence of schizophrenia on the basis of high-resolution structural MRI data. The pipeline starts off with an intensive preprocessing step that includes skull stripping, bias field correction, normalisation to MNI space, resampling to a uniform$128 \times 128 \times 128$voxel grid, and$z$-score intensity normalisation, producing standardised and clean input data. An Attention U-Net architecture is then used to divide the important regions of the brain like the prefrontal cortex, hippocampus, and thalamus as well as to apply attention gates to select such features that are of much interest in schizophrenia and eliminate irrelevant business information. To train a fine-tuned ResNet-18 model to binary classification (schizophrenia vs. control) with dropout regularization and gradual unfreezing strategy to optimize learning is then trained using these anatomically segmented outputs. The decisions of the model are also explained with the help of Grad-CAM visualizations to provide an idea about important discriminative areas. The proposed method demonstrates excellent diagnostic accuracy (93.4 %), F1-score$(93.1 \%)$, and AUC-ROC (0.961) when evaluated on the OpenNeuro dataset with 176 subjects using 5-fold crossvalidation, and the result is significantly high in comparison with baseline CNN and U-Net models. This integrated assessment methodology shows both diagnostic validity and clinical elucidation and is a promising technique in objective diagnosis of schizophrenia on the basis of the image.