A multi-atlas ensemble learning framework for autism spectrum disorder diagnosis

Clinical autism spectrum disorder (ASD) relies on subjective scales and neuroimaging, with traditional approaches suffering from high subjectivity and insufficient feature extraction, increasing misdiagnoses risk.

To improve the diagnostic accuracy, we propose a multimodal ASD prediction network (APN) that integrates resting-state functional magnetic resonance imaging (rs-fMRI) with non-imaging data. Algorithmically, we first introduce a novel cumulative distribution divergence score (CDDS) feature selection method based on the multi-atlas functional connectivity, which quantifies statistical distributional differences between positive and negative samples, effectively addressing the limitations of the traditional F-score in overlapping or weakly discriminative features. Building upon this, we adapt existing deep learning paradigms to construct a deep neural architecture (VAEA), combining a variational autoencoder (VAE), a denoising autoencoder (DAE), and a multilayer perceptron (MLP). The combination of these modules is non-trivial: the novel CDDS first purifies the high-dimensional feature space, enabling the VAEA to learn robust latent probabilistic representations via unsupervised pre-training without suffering from the curse of dimensionality. This is followed by the supervised fine-tuning to optimize parameters. We further incorporate demographic factors including age, full-scale intelligence quotient (FIQ), and gender via a late-fusion strategy to prevent them from being overshadowed by imaging features, thereby enhancing the discriminability. Functional connectivity features from three atlases including AAL, CC400, and Dosenbach160 are weighted and integrated to achieve the effective multi-source fusion.

We conduct 10-fold cross-validations on the Autism Brain Imaging Data Exchange (ABIDE) dataset, and the proposed APN achieves 81.63% accuracy, 81.89% sensitivity and 81.46% specificity, significantly outperforming state-of-the-art methods. Interpretability analysis identifies hippocampus, amygdala, and cingulate gyrus as strongly ASD-associated regions, offering neural mechanistic insights.

This study provides an efficient, objective, and interpretable multimodal deep learning framework for ASD assisted diagnosis, with substantial clinical application and neuroscience value.