Extended Invariant Information Clustering Is Effective for Leave-One-Site-Out Cross-Validation in Resting State...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Imaging

Background:

• Autism spectrum disorder (ASD) diagnosis is challenging due to subject variability, especially in children.
• Multi-source datasets in neuroimaging present difficulties for consistent model training and validation.
• Leave-one-site-out cross-validation (LOSO-CV) is a robust method for evaluating model generalizability across different data sources.

Purpose of the Study:

• To introduce a novel deep neural network model, Extended Invariant Information Clustering (EIIC), for improved LOSO-CV performance.
• To enhance the classification accuracy of autism spectrum disorder using resting-state functional connectivity magnetic resonance imaging data.
• To demonstrate the effectiveness of EIIC in transfer learning without requiring data augmentation techniques.

Main Methods:

• Developed Extended Invariant Information Clustering (EIIC), a contrastive learning model.
• Applied EIIC to the Autism Brain Imaging Data Exchange (ABIDE) resting-state fMRI dataset.
• Utilized LOSO-CV to evaluate EIIC's performance across different scanning sites, adjusting mini-batch size for optimization.

Main Results:

• EIIC achieved higher LOSO-CV classification accuracy for most scanning locations compared to existing methods.
• The model demonstrated classification accuracy greater than 0.8 for sites with the highest mean subject age.
• EIIC showed superior performance over other classifiers, particularly with optimized mini-batch sizes.

Conclusions:

• The proposed EIIC model offers a promising approach for improving diagnostic accuracy in autism spectrum disorder.
• EIIC's effectiveness in LOSO-CV suggests its potential for data harmonization across different neuroimaging sites.
• The model's simplicity and flexibility make it suitable for application in other research domains requiring robust cross-source validation.