Detection of Bipolar Disorder and Schizophrenia Employing Bayesian-Optimized Grad-CAM-Driven Deep Learning

Diagnosing bipolar disorder (BD) and schizophrenia (SCH) presents significant challenges due to overlapping symptoms, reliance on subjective assessments, and the late-stage manifestation of many symptoms. Current methods using structural magnetic resonance imaging (sMRI) as input data often fail to provide the objectivity and sensitivity needed for early and accurate diagnosis. sMRI is well known to be capable of detecting anatomical changes, such as reduced gray matter volume in SCH or cortical thickness alterations in BD. However, advanced techniques are required to capture subtle neuroanatomical patterns critical for distinguishing these disorders in sMRI. Deep learning (DL) has emerged as a transformative tool in neuroimaging analysis, offering the ability to automatically extract intricate features from large datasets. Building on its success in other domains, including autism spectrum disorder and Alzheimer's disease, DL models have demonstrated the potential to detect subtle structural changes in BD and SCH. Recent advancements suggest that DL can outperform traditional statistical methods, offering higher classification accuracy and enabling the differentiation of complex psychiatric disorders. In this context, this study introduces a novel deep learning framework for distinguishing BD and SCH using sMRI data. The model is specifically designed to address subtle neuroanatomical differences, offering three key contributions: (1) a tailored DL model that leverages explainability to extract features that boost psychiatric MRI analysis performance, (2) a comprehensive evaluation of the model's performance in classifying BD and SCH using both spatial and morphological analysis together with classification metrics, and (3) detailed insights, which are derived from both quantitative (performance metrics) and qualitative analyses (visual observations), into key brain regions most relevant for differentiating these disorders. The results have achieved an accuracy of 78.84%, an area under the curve (AUC) of 83.35%, and a Matthews correlation coefficient (MCC) of 59.10% using the proposed framework. These metrics significantly outperform traditional machine learning models. Furthermore, the proposed method demonstrated superior precision and recall for both BD and SCH, with notable improvements in identifying subtle neuroanatomical patterns. Depending on the acquired result, it can be said that the proposed method enhances the application of DL in psychiatry, paving the way for more objective, non-invasive diagnostic tools with the potential to improve early detection and personalized treatment.