Premium
MO‐F‐BRA‐06: Systematic Evaluation of a Deformable Image Registration Algorithm from a Commercial Software Package
Author(s) -
Stanley N,
Zhong H,
GlideHurst C,
Chetty I,
Movsas B
Publication year - 2012
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 - 10.1118/1.4735825
Subject(s) - imaging phantom , image registration , voxel , computer science , displacement (psychology) , medical imaging , image noise , computer vision , artificial intelligence , nuclear medicine , algorithm , medicine , image (mathematics) , psychology , psychotherapist
Purpose : As adaptive radiotherapy (ART) becomes more efficient for clinical use, verifying the accuracy of deformable image registration (DIR) is necessary. This study presents a method to evaluate a commercial DIR software package systematically for ART implementation. Methods : We developed multiple computational phantoms to evaluate Velocity Advanced Imaging (VelocityAI, B‐spline‐based algorithm) software. Phantoms were developed with a finite element method (FEM), from CT images of four lung cancer patients by simulating different diaphragm deformations, and also from the CT image of a prostate patient by modeling expansion of the bladder volume. CBCT noise in the simulated prostate phantom was included using a 3D‐noise‐power spectrum. FEM‐based displacement vectors used in the phantom construction served as a gold standard to evaluate the performance of VelocityAI's algorithms at each voxel in the image volume. In addition to computational phantoms, a deformable, physical phantom was developed to validate VelocityAI using automatically tracked landmarks. Results : VelocityAI performed well in high and low‐contrast regions. Average displacement errors range from 0.8–2.6mm, depending on the patient anatomy, magnitude of displacement and target site. With the computational lung phantoms, average errors were between 0.9–2.6mm, and in the physical phantom they were between 0.8–1.9mm. In the prostate phantom, average errors were between 1.0–1.9mm in the prostate, 1.9–2.4mm in the rectum and 1.8–2.1mm across the entire patient volume. Regions where the displacement was large tended to have more registration errors, especially when they bordered rigid structures like bone. Conclusions : VelocityAI (B‐spline‐based) outperforms the Demons algorithm in regions of largely deformed, homogenous tissue. The default parameter setting of VelocityAI is nearly optimal and the CBCT noise has no significant impact on its performance. VelocityAI can be used for DIR of lung and prostate CT images when errors on the order of 2.6 mm are acceptable. The research was supported by NIH/NCI R01CA140341.