Vision transformer and complex network analysis for autism spectrum disorder classification in T1 structural MRI

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Machine Learning

Background:

• Autism spectrum disorder (ASD) diagnosis relies on subjective methods, delaying intervention.
• Early ASD diagnosis is crucial for improving long-term outcomes.
• An automated diagnostic approach using neuroimaging can enhance efficiency and accessibility.

Purpose of the Study:

• To develop and evaluate an AI-based method for automated ASD diagnosis.
• To leverage T1-weighted structural MRI (sMRI) scans for ASD classification.
• To compare the performance of different AI models, including complex network analysis (CNA) and vision transformers (ViT).

Main Methods:

• Utilized sMRI data from 79 ASD patients and 105 controls from the ABIDE database.
• Developed features using Complex Network Analysis (CNA) and Vision Transformers (ViT).
• Trained and evaluated five models (logistic regression, SVM, GB, KNN, NN) for each feature type and federated models using fivefold cross-validation.

Main Results:

• The federated model CNA(KNN)-ViT(NN) demonstrated superior performance with 0.951 accuracy and 0.980 AUC-ROC.
• ViT-based models outperformed CNA-based models in 80% of performance metrics.
• Federating CNA and ViT models improved overall diagnostic accuracy for ASD.

Conclusions:

• AI models, particularly the federated CNA(KNN)-ViT(NN), show feasibility for automated ASD diagnosis from T1 sMRI.
• The proposed method offers a potential pathway for more efficient and accessible ASD screening in clinical settings.
• Automated diagnosis using sMRI scans can complement traditional methods, aiding timely intervention.