Early autism diagnosis based on path signature and Siamese unsupervised feature compressor

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Autism spectrum disorder (ASD) is a growing public health concern.
• Early ASD diagnosis is critical for effective intervention but challenging in infants under two years old.
• Current diagnostic methods relying on behavioral patterns lack reliability in very young children.

Purpose of the Study:

• To develop a novel deep learning-based method for early autism spectrum disorder diagnosis using structural MRI.
• To address challenges of scarce, imbalanced, and heterogeneous data in infant neuroimaging.
• To extract key neurodevelopmental features for improved diagnostic accuracy.

Main Methods:

• Proposed a Siamese verification framework to augment limited data.
• Implemented an unsupervised compressor to extract salient features and mitigate data imbalance.
• Utilized weight constraints to handle sample heterogeneity and Path Signature for longitudinal analysis.
• Identified machine learning-focused brain regions relevant for ASD diagnosis.

Main Results:

• The proposed deep learning method demonstrated strong performance in practical diagnostic scenarios.
• The approach outperformed existing machine learning methods for early autism diagnosis.
• The study provided novel anatomical insights into early autism spectrum disorder development.

Conclusions:

• The novel deep learning framework offers a promising solution for reliable early autism diagnosis in infants.
• This method effectively handles data limitations inherent in infant neuroimaging studies.
• The findings contribute to advancing early detection and intervention strategies for autism spectrum disorder.