Schizophrenia (SZ) is a multifaceted neurological disorder influenced by various factors, including brain chemistry, genetics, birth complications, and delusions. Functional Magnetic Resonance Imaging (fMRI) serves as a valuable technique for SZ detection, capturing subtle changes in blood flow associated with brain activity. However, manual screening of fMRI scans for SZ is susceptible to errors, time-intensive, and complicated by image contamination. In this study, we introduce a pioneering method for SZ diagnosis that integrates level analysis, seed-based voxel activation, and adaptive statistical parametric mapping techniques. This approach supersedes traditional methods by incorporating pre-processing and voxel activation into fMRI images, enhancing intensities and enabling the identification of brain networks or regions with significant effects at a group level, thus improving model accuracy. We leverage pre-trained ImageNet dataset deep learning models (DLMs) such as VGG- 16, ResNet50, MobileNet, and a newly developed simplified DLM named SZ-Net. Our findings demonstrate that SZ-Net achieves an impressive 10-fold validation classification accuracy of 99.24%, highlighting its proficiency in accurately categorizing fMRI scans. Additionally, SZ-Net requires fewer learnable parameters compared to pre-trained models, resulting in a more efficient and compact architecture. The proposed system holds promise for enhancing the performance of medical instruments, particularly in SZ detection using fMRI images.