Encrypted login | home

Program Information

Accuracy Preserving Normalization of Deformable Image Registration Solutions for Robust CT-Derived Ventilation Imaging

E Castillo

E Castillo1*, J Zhang1 , Y Vinogradskiy2 , R Castillo3 , T Guerrero1 , (1) William Beaumont Hospital Research Institute, Royal Oak, MI, (2) University of Colorado Denver, Aurora, CO, (3) Univ Texas Medical Branch at Galveston, Galveston, TX


TU-RPM-GePD-I-5 (Tuesday, August 1, 2017) 3:45 PM - 4:15 PM Room: Imaging ePoster Lounge

Purpose: To improve the robustness and reproducibility of CT-derived ventilation (CT-V) imaging. CT-V employs deformable image registration (DIR) to calculate local tissue volume changes between inhale/exhale CT pairs. Thus, CT-V quality is dependent on DIR spatial accuracy. However, subtle differences between sub-voxel accurate DIRs can cause significant variations in the corresponding CT-V images. We propose a DIR normalization strategy that reduces these variations while preserving spatial accuracy.

Methods: Given a DIR solution, the accuracy-normalized solution (ANS) is defined as the "smoothest" (minimum gradient magnitude) displacement field satisfying the following c-magnitude difference condition: the maximum magnitude difference between the ANS displacement vectors and the original DIR displacement vectors is less than a scalar constant c. Every DIR solution has a unique ANS. If two DIR solutions satisfy a 2c-magnitude difference condition, then their ANS's are equivalent. In order to demonstrate the utility of ANS, we apply two DIR algorithms to the same inhale/exhale CT image pair and quantify the differences between their spatial accuracies, as measured by 300 landmarks point pairs, and their corresponding CT-V values.

Results: DIR spatial accuracies for Algorithms 1 & 2 were, 1.00 (0.74) and 1.14 (0.63) respectively, while the mean difference in CT-V values was 0.07 (0.07). For Algorithm 1, the ANS spatial accuracies for c = 1 and 3 were 1.35 (0.70) and 1.38 (0.71), while for Algorithm 2, they were 1.36 (0.72), and 1.38 (0.73). The mean CT-V differences were reduced to 0.02 (0.03) and 0.01 (0.01) for the c=1,3 ANS's corresponding to Algorithms 1 and 2.

Conclusion: As demonstrated by numerical experiments, the ANS approach reduces the effects of subvoxel DIR errors on the CT-V calculation, while also maintaining spatial accuracy. The results indicate that ANS reduces the variation between CT-V images computed from similarly accurate DIRs.

Contact Email: