This thesis investigates the effectiveness of five deep-learning architectures for medical image classification (Schizophrenia). The COBRE dataset was used to train and evaluate the models. The outcomes demonstrated notable differences in performance, with high accuracy rates being attained by CNN models. While the CNN-LSTM hybrid model and the enhanced CNN model demonstrated encouraging performance, hardware constraints must be taken into account. The research highlights how crucial it is to choose the right deep-learning architectures for tasks involving the classification of medical images.