Comparative evaluation of registration algorithms in different brain databases with varying difficulty: results and...

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Anatomy

Background:

• Inter-subject registration of brain magnetic resonance images (MRI) is crucial for large-scale studies.
• Existing evaluations of registration algorithms are often limited to specific tasks or datasets.
• This limits the generalizability and applicability of chosen algorithms in diverse neuroimaging research.

Purpose of the Study:

• To evaluate the generality, accuracy, and robustness of 12 general-purpose brain MRI registration algorithms.
• To determine if algorithms can perform consistently well across various tasks and databases with minimal parameter tuning.
• To address the need for reliable registration methods in multi-institutional, large-scale neuroimaging initiatives.

Main Methods:

• Evaluated 12 registration algorithms using fixed, developer-suggested parameters.
• Tested algorithms on 7 databases/tasks encompassing 4 common challenges: anatomical variability, image quality, multi-site differences, and pathology.
• Assessed registration accuracy using expert-annotated landmarks and regions of interest (ROIs) across 7,562 registrations.

Main Results:

• Algorithm performance varied significantly across different tasks and databases.
• No single algorithm demonstrated consistent top performance across all tested challenges.
• Parameter settings and algorithm characteristics (similarity metrics, transformation models) influenced registration outcomes.

Conclusions:

• The choice of brain MRI registration algorithm is highly dependent on the specific task and data characteristics.
• Current general-purpose algorithms may not be universally applicable or robust for all large-scale neuroimaging applications.
• Further development is needed for registration algorithms that offer greater generality, robustness, and require less parameter optimization.