Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition where early detection is crucial for effective intervention. This study investigates DL-based methods for autism detection and classification using structural brain magnetic resonance imaging, with a focus on both performance and interpretability. A baseline ResNet50 model was compared with four variants enhanced with attention: self-attention, convolutional block attention module (CBAM), squeeze and excitation (SE), and a fusion attention approach (FA). The results demonstrated that the FA model achieved the most stable and superior performance, with a consistent accuracy of 97%. Although other models with enhanced attention also showed strong performance ranging from 95–97%, the fusion approach maintained better convergence throughout the training. To interpret the learned representations, saliency maps were generated using Grad-CAM, Grad-CAM++, Score-CAM, and RISE. These visualizations consistently highlighted neuroanatomical regions, all of which are frequently involved in ASD. Promising results are tempered by the limited diversity of the current dataset. Future research will expand the diversity of the cohort and influence advanced architectures to improve the accuracy of autism diagnosis.