A Classification and Pathology Study on Schizophrenia Based on Self-Attention Model

Schizophrenia is a severe mental disorder characterized by symptoms such as hallucinations and delusions that significantly impair patients' daily life. Magnetic resonance imaging has provided high-quality data support for research on psychiatric disorders due to its high imaging resolution and low side effects. However, conventional medical analysis methods based on statistics require a considerable amount of manual feature engineering. Moreover, machine learning and traditional deep learning methods struggle to perform well on high-dimensional MRI data, resulting in a bottleneck in further schizophrenia research. In this study, the self-attention mechanisms are applied to design a more efficient neural network for disease classification and pathological analysis, which does not depend on conventional convolutional approach. One significant advantage of self-attention mechanisms is that they can fully extract global representations, enabling the model to effectively leverage the complex MRI data. Additionally, by extracting the self-attention maps of the model, the proposed approach can be applied to detect critical pathological brain regions related to schizophrenia. It achieved end-to-end, purely data-driven pathological analysis and significantly reduced the cost of feature engineering. Experimental results show that the average classification accuracy of the self-attention-based neural network model is 76.76%, superior to traditional classification methods. It is also revealed that that structural damage to the thalamus, cerebellum, and surrounding regions is the main cause of schizophrenia. This model provides new insights for schizophrenia research using MRI data and for understanding the disease mechanism.