Abstract
We present NeuroLingua-Bridge, a framework for site-invariant Autism Spectrum Disorder (ASD) diagnosis that maps resting-state functional MRI (fMRI) into the token space of a large language model (LLM) while remaining invariant to the acquisition site. Language-aligned fMRI representations degrade sharply across the seventeen sites of ABIDE-I, where scanner and protocol differences dominate the diagnostic signal. Our framework trains a text-aligned tokenizer jointly with three domain-generalization components, namely a multi-class K-way site-adversarial gradient-reversal layer, site-balanced dynamic latent prototypes, and episodic Group Distributionally Robust Optimization (Group-DRO) for the worst-performing site. Under a strict leave-site-out protocol on ABIDE-I across 17 sites parcellated with the CC200 atlas, NeuroLingua-Bridge reaches 80.0% accuracy, 81.1% AUC, and 74.1% worst-site accuracy, a 15.9-point improvement over a site-agnostic variant, outperforming supervised and foundation-model baselines despite a stricter evaluation protocol.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ELAMI_005.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/ELAMI_005_supp.pdf
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{ElKha_Diagnosis_MICCAISAT2026,
author = { El Khamlichi, Khalid AND Bazay, Fatima Ez-Zahraa AND El Maliani, Ahmed Drissi},
title = { { Diagnosis of Autism Spectrum Disorder using LLMs and Multimodal Brain Connectivity Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}