3D CNN for neuropsychiatry: Predicting Autism with interpretable Deep Learning applied to minimally preprocessed...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Neuropsychiatric conditions require robust, reproducible brain-based markers.
• Deep learning (DL) shows promise but faces challenges in 3D imaging and interpretability.
• Current machine learning (ML) approaches often require extensive data preprocessing.

Purpose of the Study:

• To develop an interpretable 3D deep learning pipeline for Autism diagnosis using structural MRI.
• To bypass the need for template space transformation in predictive modeling.
• To identify brain regions associated with Autism through model interpretation.

Main Methods:

• Trained 3D deep learning models on the ABIDE I and II datasets (n=1329) for Autism diagnosis.
• Applied models to minimally processed structural MRI scans without template normalization.
• Utilized model interpretation techniques to identify contributing brain regions.

Main Results:

• Achieved predictive accuracies comparable to previous ML studies.
• Successfully identified Autism-related structural alterations in a left-lateralized language network.
• Demonstrated a method that reduces preprocessing time and resource demands.

Conclusions:

• An interpretable 3D DL pipeline can accurately infer Autism diagnosis from structural MRI.
• The method offers a computationally efficient alternative to traditional ML pipelines.
• Findings highlight the potential for DL in identifying neurobiological markers for Autism and other conditions.