A reference dataset for deformable image registration spatial accuracy evaluation using the COPDgene study archive.

Area of Science:

• Medical Imaging
• Radiology
• Computational Anatomy

Background:

• Deformable Image Registration (DIR) accuracy is crucial for medical image analysis.
• Landmark point-pairs are a validated method for assessing DIR spatial accuracy.
• Publicly available datasets are essential for reproducible DIR evaluation.

Purpose of the Study:

• To augment the www.dir-lab.com database with extensive manually identified anatomical landmark pairs.
• To provide a common dataset for evaluating DIR spatial accuracy using breath-hold computed tomography (BH-CT) images.
• To assess inter- and intra-observer spatial variation in feature localization for DIR evaluation.

Main Methods:

• Selected 10 BH-CT image pairs from the COPDgene study.
• Manually identified large sets of anatomical feature pairs between inspiratory and expiratory CT images using in-house software.
• Estimated observer variation through repeat measurements by multiple observers.
• Added 7298 landmark features to the www.dir-lab.com database.

Main Results:

• Feature pair counts per case ranged from 447 to 1172.
• Average 3D Euclidean landmark displacements varied from 12.29 to 30.90 mm across cases.
• Observer localization error estimates ranged from 0.58 to 1.06 mm.
• Observer variance was consistent across 4D CT and COPDgene cohorts.

Conclusions:

• The addition of these landmark pairs significantly broadens the applicability of reference data for DIR evaluation.
• The freely available dataset facilitates critical evaluation of DIR spatial accuracy in diverse clinical settings.
• Findings suggest consistent spatial accuracy for observers across different patient cohorts.