Automatic recognition of specific local cortical folding patterns.

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Biology

Background:

• Local cortical folding patterns are linked to psychiatric illnesses and cognitive functions.
• Manual classification of these patterns is difficult, time-consuming, and unreliable due to high variability.
• Large-scale studies are needed to understand the impact of these patterns on cortical organization.

Purpose of the Study:

• To develop and evaluate automatic algorithms for classifying local cortical folding patterns.
• To enable extended and confirmed morphological studies on large databases.
• To address the challenge of classifying rare and variable sulcal patterns.

Main Methods:

• Proposed three distinct algorithms: Support Vector Machine (SVM), Scoring by Non-local Image Patch Estimator (SNIPE), and 3D Convolutional Neural Network (CNN).
• Tested the algorithms on Anterior Cingulate Cortex (ACC) patterns and the rare Power Button Sign (PBS).
• Evaluated model performance based on balanced accuracy and efficiency in handling unbalanced datasets.

Main Results:

• Achieved approximately 80% balanced accuracy for ACC pattern classification.
• Achieved approximately 60% balanced accuracy for the rare PBS classification.
• CNN model showed rapid execution for ACC patterns; SVM and SNIPE were more effective for unbalanced data like PBS.

Conclusions:

• The proposed algorithms offer a reliable method for automatic classification of cortical folding patterns.
• These tools facilitate large-scale morphological studies, contributing to understanding brain structure-function relationships.
• The choice of algorithm depends on the specific pattern's prevalence and classification requirements.