Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{198339,
        author = {Mrs.K. Emily Esther Rani and Partheeba Pandy},
        title = {A Multimodal Deep Learning Framework for Autism Spectrum Disorder Prediction Using Structural and Functional MRI with Explainable AI},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {11},
        pages = {11106-11112},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=198339},
        abstract = {Autism Spectrum Disorder (ASD) is a prevalent neurodevelopmental condition with heterogeneous neurobiological underpinnings. While magnetic resonance imaging (MRI) has revealed structural and functional brain alterations in ASD, translating these findings into a robust, clinically actionable diagnostic aid remains a challenge. We propose an end-to-end, explainable deep learning framework that synergistically combines structural MRI (sMRI) and resting-state functional MRI (rs-fMRI) to improve ASD classification. Our system employs a two-stream convolutional neural network (CNN) architecture. The sMRI stream utilizes a 3D-ResNet to capture volumetric and morphometric features. The rs-fMRI stream first computes subject-level functional connectivity (FC) matrices, which are then processed by a CNN-Transformer hybrid model to learn both local connectivity patterns and global brain network dynamics.
The embeddings from both streams are fused via a late-fusion strategy and fed into a classification head. We leverage the Autism Brain Imaging Data Exchange (ABIDE I/II) datasets for training and evaluation. Our model outperformed unimodal benchmarks, achieving an AUROC of 0.896 compared to 0.781 (sMRI-only) and 0.823 (rs-fMRI-only). The model demonstrated superior generalizability in leave-one-site-out validation (mean AUROC: 0.861 ± 0.041) and exceeded state-of-the-art methods on a held-out test set (AUROC: 0.902). This work demonstrates the potential of integrating complementary information from sMRI and rs-fMRI within a unified, explainable deep learning pipeline to enhance the predictive accuracy and clinical interpretability of ASD biomarkers.},
        keywords = {Autism Spectrum Disorder, Deep Learning, Multimodal Fusion, sMRI, rs-fMRI, Computer-Aided Diagnosis.},
        month = {April},
        }