Premium
SU‐E‐J‐97: Quality Assurance of Deformable Image Registration Algorithms: How Realistic Should Phantoms Be?
Author(s) -
Saenz D,
Kim H,
Chen J,
Stathakis S,
Kirby N
Publication year - 2015
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4924184
Subject(s) - imaging phantom , image registration , quality assurance , artificial intelligence , image quality , image warping , computer vision , computer science , segmentation , image processing , algorithm , mathematics , nuclear medicine , image (mathematics) , medicine , external quality assessment , pathology
Purpose: Deformable image registration (DIR) has widespread uses in radiotherapy for applications such as dose accumulation studies, multi‐modality image fusion, and organ segmentation. The quality assurance (QA) of such algorithms, however, remains largely unimplemented. This work aims to determine how detailed a physical phantom needs to be to accurately perform QA of a DIR algorithm. Methods: Virtual prostate and head‐and‐neck phantoms, made from patient images, were used for this study. Both sets consist of an undeformed and deformed image pair. The images were processed to create additional image pairs with one through five homogeneous tissue levels using Otsu's method. Realistic noise was then added to each image. The DIR algorithms from MIM and Velocity (Deformable Multipass) were applied to the original phantom images and the processed ones. The resulting deformations were then compared to the known warping. A higher number of tissue levels creates more contrast in an image and enables DIR algorithms to produce more accurate results. For this reason, error (distance between predicted and known deformation) is utilized as a metric to evaluate how many levels are required for a phantom to be a realistic patient proxy. Results: For the prostate image pairs, the mean error decreased from 1–2 tissue levels and remained constant for 3+ levels. The mean error reduction was 39% and 26% for Velocity and MIM respectively. For head and neck, mean error fell similarly through 2 levels and flattened with total reduction of 16% and 49% for Velocity and MIM. For Velocity, 3+ levels produced comparable accuracy as the actual patient images, whereas MIM showed further accuracy improvement. Conclusion: The number of tissue levels needed to produce an accurate patient proxy depends on the algorithm. For Velocity, three levels were enough, whereas five was still insufficient for MIM.