Brain Biomarker Interpretation in ASD Using Deep Learning and fMRI.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Biomarker Discovery

Background:

• Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder.
• Identifying reliable biomarkers is crucial for understanding ASD's origins, enabling earlier diagnosis and personalized treatments.
• While deep neural networks (DNNs) show promise in analyzing functional magnetic resonance imaging (fMRI) for ASD identification, their decision-making processes remain underexplored.

Purpose of the Study:

• To develop and interpret reliable biomarkers for identifying ASD using fMRI data.
• To propose a novel 2-stage method for classifying ASD and control subjects and interpreting the saliency features identified by a DNN classifier.
• To address the gap in understanding the computational decision-making procedures of DNNs in ASD research.

Main Methods:

• A two-stage approach was employed, beginning with training an accurate DNN classifier on fMRI images.
• A novel frequency-normalized sampling method was developed to corrupt fMRI images for biomarker detection, leveraging brain anatomical structures.
• A new method was introduced to detect and categorize important brain features into three types for ASD versus control subject classification.

Main Results:

• The proposed method successfully identified robust and consistent biomarkers for ASD, aligning with existing literature.
• The identified biomarkers were validated through neurological function decoding and comparison with DNN activation maps.
• The study provides a deeper understanding of the data-driven computational decisions made by DNNs in identifying ASD.

Conclusions:

• The developed 2-stage method effectively identifies reliable ASD biomarkers from fMRI data.
• The findings contribute to a better understanding of ASD's neurobiological underpinnings.
• This approach holds potential for improving early ASD diagnosis and guiding the development of targeted therapies.