Automatic large quantity landmark pairs detection in 4DCT lung images.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Deformable image registration (DIR) is crucial for analyzing changes in lung CT images, particularly between different respiratory phases.
• Accurate evaluation of DIR algorithms relies on high-quality benchmark datasets with precise landmark correspondences.
• Current lung CT benchmark datasets often lack sufficient quantity and positional accuracy of landmark pairs.

Purpose of the Study:

• To develop an automated method for precise detection of a large number of landmark pairs between lung CT images.
• To support and enhance the evaluation of deformable image registration (DIR) algorithms.
• To augment existing lung CT benchmark datasets with accurate landmark pairs.

Main Methods:

• Utilized Harris-Stephens corner detection on lung vasculature probability maps to identify initial landmarks.
• Employed a parametric image registration method (pTVreg) for initial correspondence between end-exhalation (EE) and end-inhalation (EI) phases.
• Developed a multi-stream pseudo-siamese (MSPS) network to refine landmark positional accuracy by predicting 3D shifts.

Main Results:

• Successfully detected an average of 1886 landmark pairs per case across 10 4DCT lung datasets.
• Achieved a mean Target Registration Error (TRE) of 0.47 ± 0.45 mm on digital phantoms, with 98% of pairs < 2 mm.
• Obtained a TRE of 0.73 ± 0.53 mm on the DIRLAB benchmark, with 97% of pairs < 2 mm.

Conclusions:

• A new, automated method enables precise detection of numerous landmark pairs in lung CT image pairs.
• The generated landmark pairs serve as valuable benchmark datasets for quantitative evaluation of DIR algorithms.
• This approach facilitates more accurate and informative assessments of DIR performance in medical imaging.