The current clinical diagnosis of psychiatric disorders relies heavily on subjective assessment of symptoms. While neuroimaging has made an essential contribution to characterizing the brain of psychiatric disorders, it does not currently serve the clinical diagnosis of major psychiatric disorders. Here, we report a neuroimaging-aided diagnostic system for major psychiatric disorders designed for clinical needs. We developed novel deep learning networks with attentional mechanisms and applied them to a large-scale, single-center neuroimaging dataset containing four major psychiatric disorders and healthy groups (n=2490). Both cross-validation and extensive independent validation using multiple open-source datasets (n = 1972) showed that the system could accurately identify any one of the four diagnostic categories and healthy population from brain structural imaging. For the first time, we have constructed an automatic neuroimaging-aid diagnostic system that considers common issues in practice, such as co-morbid diagnoses and the discrimination between specific suspected diagnoses. Furthermore, real-world applications have validated the system’s effectiveness. These works contribute to the translation of brain research to objective diagnostic aids for psychiatric disorders.