Novel Deep Learning Approaches for Schizophrenia Diagnosis Using rsfMRI

Schizophrenia (SC) is a complex and debilitating mental disorder that affects millions worldwide. Traditional diagnostic methods often lack precision, necessitating the exploration of advanced neuroimaging techniques and machine learning algorithms for improved diagnostic accuracy. This study employs deep learning, specifically convolutional neural networks (CNN), to classify SC using resting-state functional magnetic resonance imaging (rsfMRI) data. The methodology involves preprocessing the rsfMRI data to normalize, denoise, and enhance contrast. We introduce a novel slice-wise classification approach, combining high-performing slices with significant time points to enhance diagnostic accuracy. This process transforms the input data dimensions and significantly increases the dataset size, capturing both spatial and temporal information. We further utilize transfer learning with pretrained models such as VGG16, ResNet50, and Xception, leveraging their superior feature extraction capabilities. These models are adapted and fine-tuned for our specific classification task. Our approach demonstrates significant improvements, achieving classification accuracies up to 98.75 %, thereby highlighting the potential of deep learning in SC diagnosis. Despite these advancements, challenges remain, including the limited availability of comprehensive rsfMRI datasets and the necessity for interpretable models to elucidate the neurobiological underpinnings of SC. Future research should focus on overcoming these limitations by utilizing large-scale multimodal datasets and developing more interpretable, robust machine learning models to advance the understanding and diagnosis of schizophrenia.