Deep Learning for Interpretable ADHD Diagnosis from fMRI: ConvLSTM with Enhanced Augmentation, Grad-CAM Explainability, and External Validation

The diagnosis of attention-deficit/hyperactivity disorder (ADHD) using resting-state functional MRI (rs-fMRI) remains difficult due to the high dimensionality of the data, short sample sizes, and interpretability issues with most deep learning models. In this work, based on a modified ConvLSTM architecture, we provide a robust and interpretable framework to extract important spatiotemporal biomarkers from rs-fMRI series. A targeted augmentation approach based on temporal jittering and regulated Gaussian noise, both tuned to the temporal dynamics of fMRI data, is presented to address data scarcity and class imbalance. With stratified evaluation, the model obtained an F1-score of 0.89, an accuracy of 0.90, and an AUC of 0.96 after undergoing extensive validation on the ADHD-200 dataset. Its generalizability was further validated on external cohorts that were completely independent, achieving perfect specificity and precision and a balanced accuracy of 83.3%. While these results are promising, the limited size and scope of the external cohort necessitate further validation in larger, multi-center studies. To address important transparency issues in medical AI, we integrated Grad-CAM explainability to visually identify brain regions influencing model predictions, supporting clinical applicability. In brief, this work advances the usage of deep learning in neuroimaging-based ADHD diagnosis by imparting a pipeline that is clinically relevant, reproducible, and interpretable. Overall, this study proposes a clinically significant, reproducible, and interpretable deep learning pipeline that improves diagnostic overall performance and addresses agreement and transparency, key conditions for real-world medical adoption of AI in neuroimaging