Autism Severity Detection

Overview

This research project explores multi-modal deep learning approaches to predict autism support needs as defined by the DSM-5 classification system. By combining neuroimaging data (T1 MRI scans) with potential genomic features, the goal is to develop more objective and accurate tools for assessing autism severity levels.

Research Motivation

Autism spectrum disorder (ASD) severity assessment currently relies heavily on behavioral observations and clinical interviews, which can be subjective and vary between practitioners. This project aims to complement clinical assessment with quantitative biomarkers derived from brain imaging and genetic data, potentially leading to:

• More consistent and reproducible severity classifications
• Earlier identification of support needs
• Better understanding of the neurobiological basis of autism
• Personalized intervention planning

Technical Approach

Neuroimaging Pre-processing Pipeline

Building a reproducible and robust pipeline for processing T1-weighted MRI scans:

• NIfTI I/O: Handling medical imaging file formats for reading and writing brain scans
• Skull-stripping: Implementing HD-BET and FSL BET for brain extraction
• Bias-field correction: Removing intensity inhomogeneities caused by scanner artifacts
• MNI152 Registration: Aligning all brain scans to a standard template space
• Resampling: Standardizing voxel dimensions across datasets
• Intensity Normalization: Applying WhiteStripe and z-score normalization methods

Feature Extraction

• Extracting embeddings from pre-trained neuroimaging models
• Investigating potential integration of genomic variant-level features
• Exploring multi-modal fusion strategies for combining imaging and genetic data

Model Development

• Designing architectures for multi-class classification (DSM-5 support levels)
• Implementing attention mechanisms for interpretable predictions
• Developing validation strategies appropriate for medical imaging applications

Technologies Used