A vessel bifurcation liver CT landmark pair dataset for evaluating deformable image registration algorithms

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Deformable image registration (DIR) algorithm evaluation is critical for clinical adoption.
• A significant gap exists in dependable DIR benchmark datasets, particularly for liver imaging.
• Existing datasets are insufficient for comprehensive DIR performance assessment.

Purpose of the Study:

• To introduce a comprehensive liver computed tomography (CT) DIR landmark dataset library.
• To enable efficient and quantitative evaluation of various DIR methods for liver CTs.
• To facilitate the development of more accurate and reliable liver image registration techniques.

Main Methods:

• Acquired 40 CT liver image pairs from public archives and institutions.
• Developed a semi-automatic procedure for landmark generation, including vessel segmentation and bifurcation detection.
• Employed two DIR methods for landmark correspondence and rigorous validation to ensure positional accuracy.

Main Results:

• Generated a dataset with an average of ~56 landmark pairs per image pair, totaling 2220 landmarks across 40 cases.
• Achieved target registration errors (TRE) of 0.37 ± 0.26 mm and 0.55 ± 0.34 mm on digital phantoms.
• Demonstrated high landmark accuracy, with 97% of pairs having TREs below 1.5 mm and distances to manual placement averaging 1.27 ± 0.79 mm.

Conclusions:

• This work presents the first large-scale liver CT DIR landmark dataset derived from real patient images.
• The dataset serves as a valuable ground truth resource for quantitative evaluation of DIR algorithms in liver applications.
• It is expected to significantly advance the field of liver image registration and its clinical utility.