Performance scaling for structural MRI surface parcellations: a machine learning analysis in the ABCD Study.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Developmental Neuroscience

Background:

• Predefined brain parcellations are commonly used for data reduction in neuroimaging.
• Parcellation choice is often overlooked in prediction-based studies, despite its potential impact on performance.

Purpose of the Study:

• To investigate the relationship between different brain parcellations and predictive performance.
• To assess the influence of parcellation resolution, type, and machine learning strategies on prediction accuracy.

Main Methods:

• Utilized structural magnetic resonance imaging (sMRI) data from the Adolescent Brain Cognitive Development Study (N=9,432).
• Examined 220 parcellations across 45 phenotypic measures in 9- to 10-year-old children.
• Assessed various machine learning (ML) pipelines and multi-parcellation strategies.

Main Results:

• Parcellation performance scaled with spatial resolution; more parcels generally led to better predictions (power-law scaling between 1/4 and 1/3).
• Literature-based parcellations, support vector machine (SVM) pipelines, and ensembling multiple parcellations showed superior performance.
• Higher resolution parcellations demonstrated significant performance improvements.

Conclusions:

• The selection of brain parcellation is a critical factor influencing predictive model performance.
• Optimizing parcellation resolution and strategy can substantially enhance prediction accuracy in neurodevelopmental studies.
• Findings underscore the importance of considering parcellation characteristics for robust neuroimaging-based predictions.