TRACE-Net: Trustworthy Multi-View Evidential Learning for Autism Detection from Multi-Site Functional MRI

Abstract

Reliable computer-aided screening for Autism Spectrum Disorder (ASD) from functional MRI (fMRI) faces three coupled difficulties: a single connectivity summary keeps only part of the signal, softmax outputs are over-confident and raise no flag on hard cases, and inter-scanner variation across sites weakens transfer. This paper introduces TRACE-Net, which tackles the three at once. A tri-path encoder pairs a structured state-space branch with channel self-attention and convolution to represent static and dynamic functional connectivity and regional activity descriptors, alongside a phenotypic-metadata view. Each view is turned into a Dirichlet opinion, scaled by a per-sample reliability estimate, and combined by a conflict-aware Dempster-Shafer rule, so a prediction arrives with a calibrated confidence. A gradient-reversal branch strips centre-specific cues. Evaluated on the 871-subject ABIDE-I split under 10-fold cross-validation, TRACE-Net attains 96.31% accuracy and 0.9842 AUC with better calibration than softmax and Monte-Carlo dropout, and an imaging-only version still leads every method compared.