Explainable AI for Autism Diagnosis: Identifying Critical Brain Regions Using fMRI Data

TL;DR

This work tackles the need for objective and explainable ASD diagnostics by developing a high-accuracy DL framework that classifies ASD from resting-state fMRI while providing interpretable biomarkers. The approach uses a two-stage training pipeline (unsupervised pre-training of a Stacked Sparse Autoencoder followed by supervised fine-tuning) on ABIDE data, with feature reduction via SVM-RFE and a comprehensive ROAR-based interpretability benchmark. Across multiple preprocessing pipelines, the model achieves up to 98.2% accuracy and identifies consistent visual-processing regions (e.g., calcarine sulcus, cuneus) as key discriminators, aligning with established ASD literature and supporting neurobiological validity. The study advances explainable AI in neuroimaging by benchmarking multiple interpretability methods, validating gradient-based approaches, and demonstrating robust cross-pipeline biomarker localization with potential for earlier, reliable ASD diagnosis.

Abstract

Early diagnosis and intervention for Autism Spectrum Disorder (ASD) has been shown to significantly improve the quality of life of autistic individuals. However, diagnostics methods for ASD rely on assessments based on clinical presentation that are prone to bias and can be challenging to arrive at an early diagnosis. There is a need for objective biomarkers of ASD which can help improve diagnostic accuracy. Deep learning (DL) has achieved outstanding performance in diagnosing diseases and conditions from medical imaging data. Extensive research has been conducted on creating models that classify ASD using resting-state functional Magnetic Resonance Imaging (fMRI) data. However, existing models lack interpretability. This research aims to improve the accuracy and interpretability of ASD diagnosis by creating a DL model that can not only accurately classify ASD but also provide explainable insights into its working. The dataset used is a preprocessed version of the Autism Brain Imaging Data Exchange (ABIDE) with 884 samples. Our findings show a model that can accurately classify ASD and highlight critical brain regions differing between ASD and typical controls, with potential implications for early diagnosis and understanding of the neural basis of ASD. These findings are validated by studies in the literature that use different datasets and modalities, confirming that the model actually learned characteristics of ASD and not just the dataset. This study advances the field of explainable AI in medical imaging by providing a robust and interpretable model, thereby contributing to a future with objective and reliable ASD diagnostics.

Figures (6)

• Figure 1: The overall architecture of the model, incorporating both the SSAE and softmax classifier.
• Figure 2: Loss graphs during first K-Fold. AE1, AE2, and classifier loss represent reconstruction loss during unsupervised pre-training. Model loss and accuracy refer to the supervised fine-tuning stage.
• Figure 3: Performance of the model across different preprocessing pipelines averaged across 5 K-folds.
• Figure 4: Important ROIs (yellow) identified by the Integrated Gradients approach when using DPARSF pipeline.
• Figure 5: ROAR Analysis, shows how removal of top 1% of features leads to steeper accuracy drop for the Integrated Gradients, DeepLift, DeepLiftShap and GradientShap methods.