Attention-enhanced 3D CNNs for ASD classification from structural MRI : a comparison with transformer-based models

Tools

Alkhaibari, Sahar and Dong, Feng; Ni, Hao and Cafolla, Daniele, eds. (2026) Attention-enhanced 3D CNNs for ASD classification from structural MRI : a comparison with transformer-based models. In: Artificial Intelligence in Healthcare. Lecture Notes in Computer Science (1). Springer, GBR, 381–394. ISBN 9783032353900 (https://doi.org/10.1007/978-3-032-35390-0_27)

[thumbnail of Alkhaibari-Dong-AIiH-2026-Attention-enhanced-3D-CNNs-for-ASD-classification-from-structural-MRI]
Preview
 Text. Filename: Alkhaibari-Dong-AIiH-2026-Attention-enhanced-3D-CNNs-for-ASD-classification-from-structural-MRI.pdf
Accepted Author Manuscript
License: Creative Commons Attribution 4.0 logo
Download (1MB)| Preview

Abstract

Timely identification of Autism Spectrum Disorder (ASD) is of considerable clinical importance, as early intervention has been consistently associated with improved long-term outcomes. However, current diagnostic approaches rely largely on behavioural observation and clinician-based assessments, which are often subjective and time-consuming. In this study, we design and systematically evaluate four 3D deep learning architectures of progressive complexity, ranging from a baseline CNN to a full Swin Transformer, for ASD classification from structural MRI using the ABIDE dataset (N = 98), incorporating site embeddings to account for multi-site variability. Experimental results demonstrate that incorporating lightweight channel attention substantially improves classification performance. The attention-enhanced 3D CNN achieved the highest accuracy of 85% and an F1-score of 88%, outperforming both the baseline CNN (60% accuracy, F1 = 75%) and the Swin Transformer, which offered no additional benefit under limited data conditions. These findings suggest that lightweight channel attention offers a favourable accuracy–efficiency tradeoff for ASD classification from structural MRI, representing a promising direction for automated ASD classification under limited data conditions, whereas complex Transformer architectures may not provide additional discriminative benefit in such settings.

CORE (COnnecting REpositories)