Site-Invariant Meta-Modulation Learning for Multisite Autism Spectrum Disorders Diagnosis

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Informatics

Background:

• Multisite resting-state functional magnetic resonance imaging (rs-fMRI) data is crucial for generalized brain disease prediction models.
• Site variation in multisite rs-fMRI data causes heterogeneity, hindering biomarker identification and diagnostic accuracy.
• Existing methods for multisite data face limitations in processing new sites or require extensive retraining.

Purpose of the Study:

• To develop a novel framework for brain disease diagnosis that overcomes the multisite problem in rs-fMRI data.
• To create a universal model capable of generalizing to unseen sites without requiring fine-tuning.
• To adaptively calibrate site-specific features into site-invariant features for robust diagnostic models.

Main Methods:

• A novel framework employing a learning-to-learn strategy to calibrate features.
• A modulation mechanism designed to extract site-invariant features from multisite rs-fMRI data.
• Validation using the Autism Brain Imaging Data Exchange (ABIDE I and II) dataset.

Main Results:

• The proposed framework effectively alleviates the multisite problem by converting site-specific features to site-invariant ones.
• The model demonstrated strong generalization ability, improving diagnostic accuracy on both seen and unseen multisite samples.
• The approach allows direct application to samples from unseen sites without the need for fine-tuning.

Conclusions:

• The novel framework successfully addresses the multisite challenge in rs-fMRI data for brain disease diagnosis.
• The method enhances the generalizability of diagnostic models, leading to improved accuracy across diverse data sources.
• This approach offers a practical solution for building robust, universally applicable brain disease diagnostic tools.