Deep Transfer Learning for Schizophrenia Detection Using Brain MRI

Schizophrenia causes hallucinations, delusions, and excessive disorganization. Early diagnosis and treatment lessen family issues and social expenses. It needs multiple psychiatrist consultations and brain MRIs to diagnose. Schizophrenia has no objective medical index. Machine learning and deep learning algorithms simplify disease diagnosis. Handcrafted features require domain specialists to develop the feature set, which involves time and experience in machine learning models. Deep Learning (DL) models employ brain MRI scans to predict schizophrenia; however, they require a large dataset to train, which increases computing time. But, only a few schizophrenia-detectingMRI datasets are openly available. Previously, transfer learning-based DL models have been trained on sagittal, coronal, and axial 3D MRI images. But in this model, unnecessary information and noisy features reduce performance. Therefore, we employ the axial viewof brain scans as it contains the subcortical region and ventricular areaswhich contribute most to the prediction of schizophrenia. Axial view images are used to train transfer learning-based VGG19 model for schizophrenia identification. This study uses a COBRE-published brain MRI dataset. The collection includes 146 patients' brain MRIs. We analyzed 3D MRI images to produce axial brain slices. Thresholding improves intensity differentiation in axial viewimages and data augmentation reduces overfitting and reduces data; both of them are being used in preprocessing. The suggested model utilizes a VGG-19 pre-trained network with fully connected layers. Adam optimizer for optimization, ReLU for hidden layers, and sigmoid for last layer activation functions. Our model was evaluated using accuracy, precision, recall, and F1 score. The dataset modelwas 90.9% accurate. It is compared using standard metrics to different categorization models and existing models. We improved accuracy by a maximum of 3% and a minimum of 1%.