Encrypted login | home

Program Information

Displacement Vector Field (DVF) Error Correction Using Sparsely Distributed Ground-Truth Displacement Vectors

no image available
X Liang

X Liang*, L Hu , F Yin , J Cai , Duke University Medical Center, Durham, NC

Presentations

WE-AB-BRA-3 (Wednesday, July 15, 2015) 7:30 AM - 9:30 AM Room: Ballroom A


Purpose: This study aims to develop a technique capable of reducing errors in deformable image registration (DIR) by integrating information of sparsely distributed ground-truth displacement vectors into a dense DVF.

Methods: An in-house developed lung motion phantom consisting of an end-of-inhalation (EOI) MR image, a dense DVF derived from tagging MR data, and an end-of-exhalation (EOE) image obtained by deforming the EOI image with the DVF was used. The DVFs to be corrected were obtained by registration between the EOI and the calculated EOE images using Velocity, MIM, ILK and OHS algorithms in DIRART toolbox, and Elastix. A set of sparsely distributed ground-truth points were selected and displacement vectors at those points were treated as ground-truth displacement vectors. Displacement vectors in the DVFs to be corrected were converted into coefficient vectors using Wendland's compactly supported functions. The coefficient map was modified such that the displacement vectors at ground-truth points, determined by the modified map, match the ground-truth displacement vectors. DVFs before and after correction were compared with the ground-truth DVF (phantom DVF), respectively. Pre- and post-correction registration errors were calculated.

Results: Reduced registration errors were found in all tested DVFs. Pre- and post-correction 3D errors per voxel for Velocity, MIM, ILK, OHS, and Elastix are 1.27mm vs 1.07mm, 1.56mm vs 1.55mm, 6.33mm vs 4.81mm, 3.53mm vs 2.94mm, 2mm vs 1.67mm. Reduced registration errors are also uniformly found in every direction (p<0.01).

Conclusion: We have implemented a hybrid registration technique that can potentially reduce registration errors by introducing information of sparsely distributed ground-truth displacement vectors into a dense DVF. Simulation results show that ground-truth points less than 0.5% of total number of voxels can enhance the accuracy of DVFs. Further studies on integration of biomechanical models and fine-tuning of parameters are expected to provide further correction.


Contact Email: