Transfer Learning with Transformer-Based Models and Explainable AI for Autism Detection Using Brain MRI

Background/Objectives: Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition that remains challenging to diagnose using traditional clinical methods. Recent advances in artificial intelligence, particularly transformer-based deep learning models, have shown considerable potential for improving diagnostic accuracy in neuroimaging applications. This study aims to develop and evaluate a transformer-based framework for automated ASD detection using structural and functional brain MRI data. Methods: We developed a deep learning framework using neuroimaging data comprising structural MRI (sMRI) and functional MRI (fMRI) from the ABIDE repository. Each modality was analyzed independently through transfer learning with transformer-based pretrained architectures, including Vision Transformer (ViT), MaxViT, and Transformer-in-Transformer (TNT), fine-tuned for binary classification between ASD and typically developing controls. Data augmentation techniques were applied to address the limited sample size. Model interpretability was achieved using SHAP to identify the influential brain regions that contribute to classification decisions. Results: Our approach significantly outperformed traditional CNN-based methods and state-of-the-art baseline approaches across all evaluation metrics. The MaxViT model achieved the highest performance on sMRI (98.51% accuracy and F1-score), while both TNT and MaxViT reached 98.42% accuracy and F1-score on fMRI. SHAP analysis provided clinically relevant insights into brain regions most associated with ASD classification. Conclusions: These results demonstrate that transformer-based models, coupled with explainable AI, can deliver accurate and interpretable ASD detection from neuroimaging data. These findings highlight the potential of explainable DL frameworks to assist clinicians in diagnosing ASD and provide valuable insights into associated brain abnormalities.