A Siamese Network With Node Convolution for Individualized Predictions Based on Connectivity Maps Extracted From...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning shows promise for diagnosing neuropsychiatric disorders using neuroimaging, particularly resting-state functional magnetic resonance imaging (RS-fMRI).
• A significant challenge in training deep learning models for this purpose is the limited availability of large sample sizes.

Purpose of the Study:

• To propose a novel deep learning model, the Siamese network with node convolution (SNNC), for individualized predictions using RS-fMRI data.
• To address the bottleneck of insufficient sample size in deep model training for neuroimaging analysis.

Main Methods:

• Developed a Siamese network architecture (SNNC) that utilizes sample pairs as input to mitigate issues related to small sample sizes.
• Adapted node convolution for connectivity maps derived from RS-fMRI data within the Siamese network's branches.
• Modified the loss function to mean squared error for quantitative regression, enabling prediction of label differences and individual trait estimation.

Main Results:

• The SNNC model demonstrated effective predictive performance for age and IQ on the Cam-CAN dataset, even with a minimal sample size of 40.
• SNNC achieved state-of-the-art accuracy compared to various deep learning and standard machine learning methods.

Conclusions:

• The proposed SNNC model offers a viable solution for individualized predictions from RS-fMRI data, particularly in scenarios with limited sample sizes.
• SNNC represents a significant advancement in applying deep learning to neuroimaging for objective diagnosis and trait prediction in neuropsychiatric disorders.