A comprehensive lung CT landmark pair dataset for evaluating deformable image registration algorithms.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Radiology

Background:

• Deformable image registration (DIR) is crucial for medical tasks but lacks robust validation datasets.
• Current lung CT benchmark datasets have limitations in landmark pair quantity and distribution.
• High-quality datasets are essential for advancing DIR algorithm accuracy and reliability.

Purpose of the Study:

• To develop and introduce a comprehensive lung CT deformable image registration (DIR) benchmark dataset library.
• To address the limitations of existing datasets by increasing the number and improving the distribution of landmark pairs.
• To provide a valuable resource for the quantitative validation of DIR algorithms in thoracic imaging.

Main Methods:

• Acquired 30 CT image pairs from public repositories and institutional sources.
• Developed an automated workflow for denoising, segmentation of lungs, airways, and vessels, and detection of vessel bifurcations.
• Utilized DIR to project landmarks and manually verified landmark pairs, resulting in an average of 1262 pairs per image pair.
• Estimated target registration error (TRE) using digital phantoms, achieving 0.4 mm ± 0.3 mm.

Main Results:

• Generated a lung CT DIR benchmark dataset library with an average of 1262 landmark pairs per image pair.
• Achieved a low target registration error (TRE) of 0.4 mm ± 0.3 mm, indicating high accuracy.
• The dataset is the largest of its kind, significantly enhancing the available ground truth for DIR validation.

Conclusions:

• The developed dataset library represents the most extensive collection for lung CT DIR validation to date.
• This resource will enable researchers to quantitatively assess and improve DIR algorithms for lung imaging applications.
• The dataset facilitates more robust and accurate deformable image registration in clinical practice.