Program Information
A Method to Automatically and Accurately Generate Large Number of Ground Truth Landmark Points to Verify Deformable Image Registration On Abdominal 4DCT Images
D Yang1*, Y Fu1 , H Li1 , Y Duan2 , (1) Washington University in St Louis, St Louis, MO, (2) University of Missouri, Columbia, MO
Presentations
WE-RAM1-GePD-IT-3 (Wednesday, August 2, 2017) 9:30 AM - 10:00 AM Room: Imaging ePoster Theater
Purpose: It is difficult to verify deformable image registration (DIR) results on the patient images because deformation ground truth is not available. Manually selected landmarks have been the golden standard. Though publically available for thoracic 4DCT, ground truth landmark datasets are not available for abdominal 4DCT due to the difficulties of manually and confidently determining the landmark points on low-contrast soft tissues. The purpose of this work is to develop an image processing procedure to automatically and accurately detect large quantity of landmarks in pairs on the abdominal 4DCT images.
Methods: To maximize the quantity of the detected feature points, a 3D scale-invariant feature transform (SIFT) feature detection method and a Harris-Laplacian corner detection method were employed. To allow accurate point matching, the detection was performed in a Gaussian pyramid multi-resolution scheme. For a given resolution, a novel guided-matching algorithm was developed to use higher-confidence pairs detected at both the lower resolution and the current resolution to guide the matching of additional point pairs. A Random sample consensus (RANSAC) method was applied during the matching process to detect and reject the outliers. Three patient datasets were applied for evaluation. Manual verification was performed using randomly selected 10% of the pairs.
Results: The point detection and guided-matching algorithms were optimized for abdominal 4DCT images. Greater than 9600 landmark pairs were successfully matched for each dataset, which spanned across the entire imaging volume. Matching accuracies were greater than 99% for all pairs and virtually 100% for the top 5000 pairs.
Conclusion: An automated procedure was developed to detect large quantity of landmarks in abdominal 4DCT image pairs, which allows the build of ground truth datasets to support quantitative verification of DIR algorithms on abdominal 4DCT images.
Contact Email: