ROC-based assessments of 3D cortical surface-matching algorithms

Area of Science:

• Neuroimaging analysis
• Computational anatomy
• Medical image processing

Background:

• Semi-automated brain surface analysis algorithms are increasingly used but lack rigorous performance assessment.
• Understanding algorithm performance is crucial for reliable neuroimaging studies.
• Existing methods often fail to account for sources of variance that degrade performance.

Purpose of the Study:

• To develop and present a quantitative method for assessing the performance of brain surface analysis algorithms.
• To evaluate and compare different surface-matching algorithms, including fluid flow (FF), geodesic (GD), and nearest neighbor (NN) methods.
• To identify and isolate sources of variance impacting algorithm performance in MRI datasets.

Main Methods:

• Utilized Receiver Operating Characteristic (ROC) curves to assess algorithm sensitivity and specificity in detecting synthetic deformations.
• Developed a method to isolate sources of variance (registration errors, deformation placement, morphological variability) affecting algorithm performance.
• Applied these assessment methods to compare FF, GD, and NN surface-matching algorithms on MRI data.

Main Results:

• Algorithm performance is significantly influenced by inter-individual and between-group variability in cortical surface morphology.
• The fluid flow (FF) algorithm generally performed as well as or better than the geodesic (GD) and nearest neighbor (NN) algorithms.
• GD and NN algorithms showed high variance in point correspondences and were prone to false-positive detections at high-curvature areas.

Conclusions:

• Rigorous quantitative assessment is essential for validating brain surface analysis algorithms.
• The FF algorithm demonstrates superior robustness and accuracy compared to GD and NN methods for detecting cortical deformations.
• Understanding sources of variance is key to improving the reliability and interpretability of neuroimaging analyses.