Room: Room 202
Purpose: This is a following-up study of authors â€“ automatic detection of landmark pairs (MedPhys, 12/2017). A new image processing procedure was developed to refine positions of landmark pairs that were automatically detected between intra-patient volumetric image pairs. The aim is to improve landmark pair positional accuracy so that landmark pairs can be applied for image registration verification with improved confidence.
Methods: Landmark pairs were automatically detected using the MRICGM (Multiple-Resolution Inverse-Consistency Guided Matching) method. For each landmark pair, position of the first landmark is iteratively refined by maximizing image similarity between the first block (centered at the landmark in the first image) and the second block (in the second image). A landmark pair will be flagged if the landmark repositioning distance is in the top 10% among all landmark pairs or the inverse consistency error is in the top 10%, or optimization does not converge. After all landmark pairs are processed, the flagged landmark pairs are excluded.
Results: The proposed method was implemented in MATLAB. Image intensity SSD was chosen as the similarity metric because it provided the best accuracy and computational performance among the tested similarity metrics in preliminary studies. After the flagged pairs were excluded, the refined landmark positions of the remaining landmark pairs were verified against the ground truth positions for 7 digital phantom cases. On average, ~14% of the automatically detected landmark pairs were flagged and rejected, the TRE (the distance from the landmark position to the known ground truth position) was reduced by 57%, from 1.07Â±0.94 mm to 0.46Â±0.43 mm. Also tested on the 10 DIRLAB 4DCT datasets, the proposed method was able to improve positional accuracy for 60% of the manually labelled landmarks.
Conclusion: A new procedure was developed to improve positional accuracy and overall quality of automatically detected landmark pairs.
Funding Support, Disclosures, and Conflict of Interest: The project described was partially supported by the AHRQ (Agency for Healthcare Research and Quality) grant number 1 R01 HS022888-01 and its contents are solely the responsibility of the authors and do not necessarily represent the official views of the Agency for Healthcare Research and Quality.
Image Analysis, Image Processing, Image-guided Therapy