Construction of a small-sample brain imaging data augmentation and explainable diagnostic model for autism based on...

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized primarily by social communication deficits and repetitive stereotyped behaviors. Its objective diagnosis has long relied on clinical scale assessments, lacking automated tools based on brain imaging.

This study proposes an ASD auxiliary diagnostic framework integrating conditional generative adversarial network (conditional GAN, cGAN) data augmentation, multimodal feature fusion, and explainable deep learning, based on the ABIDE I/II multi-center public datasets. First, functional connectivity matrices of AAL-116 brain regions were extracted from resting-state functional magnetic resonance imaging (rs-fMRI), and cortical morphological features were derived from structural magnetic resonance imaging (sMRI). Multi-site scanning biases were corrected using the ComBat method. On this basis, minority class samples were augmented using class-conditional GAN, followed by multimodal information fusion via a dual-branch encoder and cross-attention mechanism, ultimately outputting classification decisions between ASD and typically developing (TD) subjects.

Experimental results demonstrate that, under stratified five-fold cross-validation, the proposed method achieved an AUC of 0.871 ± 0.016 and a balanced accuracy of 0.797 ± 0.012 on the full multimodal sample set, representing improvements of 13.2% and 9.2% over single sMRI and rs-fMRI modalities, respectively. Leave-one-site-out (LOSO) cross-validation yielded an average AUC of 0.783 ± 0.041, validating the model's cross-center generalization capability.

Explainability analysis based on Integrated Gradients revealed that the default mode network and social brain regions are key decision bases for distinguishing ASD from TD, highly consistent with existing neurobiological evidence.

Not applicable.