Enhancing Autism Spectrum Disorder identification in multi-site MRI imaging: A multi-head cross-attention and...

Area of Science:

• Neuroimaging
• Machine Learning
• Developmental Disorders

Background:

• Multi-site MRI data presents significant variability due to differences in scanners and acquisition sites, complicating image analysis and model generalizability.
• Accurate and early diagnosis of Autism Spectrum Disorder (ASD) is crucial for intervention, but traditional clinical observation is time-consuming.
• Functional MRI (fMRI) combined with machine learning offers potential for expedited ASD diagnosis, yet existing methods struggle with data from diverse sources.

Purpose of the Study:

• To develop a robust neuroimaging analysis framework that minimizes site and scanner variability in multi-site datasets.
• To create a generalized machine learning model for Autism Spectrum Disorder (ASD) diagnosis that performs effectively across different MRI scanners and sites.
• To enhance the diagnostic accuracy and generalizability of fMRI-based ASD detection methods.

Main Methods:

• Proposed a Cross-Combination Multi-Scale Multi-Context Framework (CCMSMCF) for neuroimaging-based diagnostic classification.
• Integrated two novel sub-modules: the Multi-Head Attention Cross-Scale Module (MHACSM) and the Residual Multi-Context Module (RMCN).
• Employed a combination of Binary Cross Entropy, Dice loss, and Embedding Coupling loss for model training.

Main Results:

• The CCMSMCF demonstrated a degree of internal data harmonization, achieving site and scanner-agnostic performance.
• Validated the model on the multi-site Autism Brain Imaging Data Exchange I (ABIDE-I) dataset.
• Achieved promising results in diagnostic classification for Autism Spectrum Disorder (ASD) using multi-site fMRI data.

Conclusions:

• The proposed CCMSMCF framework effectively addresses the challenge of site and scanner variability in multi-site neuroimaging studies.
• This approach offers a more generalizable and robust method for machine learning-based diagnosis of Autism Spectrum Disorder (ASD).
• The CCMSMCF framework holds potential for improving the clinical application of fMRI in diagnosing developmental disorders across diverse data sources.