Automatic Individual Cortical Parcellation for the Human Connectome Project.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning in Neuroscience

Background:

• The Human Connectome Project multimodal cortical parcellation (HCP_MMP1.0) aimed to map individual human cerebral cortex using structural and functional MRI data.
• Existing methods struggled to capture individual variability in brain organization, especially for atypical cortical topologies.
• The choice of resting state network (RSN) feature maps and classification criteria for individual parcellation remained unclear.

Purpose of the Study:

• To develop an improved automated classifier for subject-specific human cortical parcellation that accounts for individual variability.
• To compare different RSN decomposition methods for their effectiveness in areal classification.
• To identify novel atypical organizations of language-related area 55b.

Main Methods:

• Introduction of the A real R ecognition E nsemble with N ested A pproach (ARENA) classifier, utilizing weakly supervised learning to handle uncertain labels.
• Comparison of temporal ICA and PROFUMO RSN maps against the original spatial ICA-based approach.
• Evaluation of classifier performance using reliable individual variability scores, areal detection rates, and test-retest reproducibility.

Main Results:

• The ARENA classifier achieved a reliable individual variability score of 8380, an average areal detection rate of 97.8%, and test-retest reproducibility of 73.3%.
• ARENA outperformed a retrained Multi-layer Perceptron (MLP) model on the same dataset and demonstrated stronger generalization to task fMRI data.
• New types of atypical organization in language-related area 55b were identified across 1071 HCP-Young Adult subjects.

Conclusions:

• The ARENA classifier significantly enhances the accuracy and reliability of individual cortical parcellation, effectively capturing individual brain organization variability.
• The study provides a data-driven HCP_MMP1.0_1071_MPM group parcellation and insights into area 55b organization.
• The developed automated individual parcellation pipeline is integrated into HCP pipelines, offering a valuable tool for neuroimaging research.