Multi-Architecture Convolutional Neural Networks with Attention Mechanisms for Autism Spectrum Disorder Classification

Background: The early identification of individuals with autism spectrum disorder (ASD) is crucial for their proper integration into the educational system and society. AI methods introduce novel approaches for the detection and classification of individuals diagnosed with ASD. Methods: We employed three custom-built Convolutional Neural Networks (CNNs) alongside two pretrained CNNs, specifically YOLO8 and ResNet18. The integrated Convolutional Block Attention Module (CBAM) was utilized to enhance feature representations for classifying individuals with ASD and non-ASD. Results: The results from the binary classification using the YOLO8-CBAM model demonstrated notable performance metrics: an accuracy of 77.6%, an F1-score of 72.6%, a Matthews Correlation Coefficient (MCC) of 60%, and an area under the curve (AUC) of 0.912. Conclusion: The backbone of the pretrained YOLO8-CBAM, enhanced by the integration of the CBAM after selected convolutional blocks, improved the feature refinement utilized in the classification process. Additionally, the gradient-weighted Class Activation Mapping (Grad-CAM) model provides interpretability by highlighting the regions that are most influential in distinguishing between individuals with ASD and those without.