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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder with primary challenges in social communication and repetitive behaviors.
• Current ASD diagnosis relies on clinical assessments, lacking automated, objective tools, particularly from brain imaging data.

Purpose of the Study:

• To develop an automated auxiliary diagnostic framework for ASD using multimodal brain imaging data.
• To enhance diagnostic accuracy and generalization capabilities through advanced AI techniques.

Main Methods:

• Utilized resting-state functional magnetic resonance imaging (rs-fMRI) and structural magnetic resonance imaging (sMRI) data from the ABIDE I/II datasets.
• Applied conditional generative adversarial networks (cGAN) for data augmentation and multimodal feature fusion via a dual-branch encoder with cross-attention.
• Corrected multi-site scanning biases using the ComBat method and employed explainable deep learning for analysis.

Main Results:

• The proposed framework achieved an Area Under the Curve (AUC) of 0.871 and balanced accuracy of 0.797 on multimodal data, outperforming single-modality approaches.
• Leave-one-site-out cross-validation demonstrated strong cross-center generalization with an average AUC of 0.783.
• The model showed significant improvements over single sMRI (13.2%) and rs-fMRI (9.2%) modalities.

Conclusions:

• Explainability analysis identified the default mode network and social brain regions as critical for distinguishing ASD from typically developing individuals.
• The findings align with existing neurobiological evidence, supporting the framework's clinical relevance.
• The study highlights the potential of integrated, explainable AI in advancing objective ASD diagnosis.