Click here to

Session: Cone-Beam CT [Return to Session]

Introduction of Lambda: A Convenient and Robust Subvoxel Image Similarity Metric

P Boyle*, M Lauria, B Stiehl, L Naumann, A Santhanam, D Low, UCLA, Los Angeles, CA


SU-H300-IePD-F8-2 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 8

Purpose: Present and characterize an intravoxel image similarity metric that incorporates physical constraints for use in CT image analysis.

Methods: Simulated volumes were produced using discretized volume averaging to approximate typical lung blood vessel profiles. Each volume measured 21x21x100 mm² and consisted of a simulated vessel surrounded by simulated parenchymal tissue with Hounsfield Unit (HU) values of 100 HU and -900 HU, respectively. Physical volume averaging was approximated using a 0.8 mm isotropic Gaussian filter. The radii of the vessels were 1, 3, or 5 mm. Each reference vessel was centered at (X, Y) = (11, 11). A corresponding set of comparison volumes were horizontally offset from the reference, from -1.0 mm to 1.0 mm in 0.05 mm steps. Each volume had 20, 40, or 60 HU Gaussian noise injected and was smoothed using a 0.5, 1.0, or a 1.5 mm isotropic Gaussian filter. Each reference-comparison pair was analyzed using λ, based on the dose-distribution comparison metric γ, with input parameters of dHU=30.0 HU and DTA=0.5 mm. The spatial sensitivity of λ was determined by inspecting the impact of the offset magnitudes on the λ distributions. The λ angle, θ, distributions were investigated as region-of-interest criteria by comparison to gradient methods.

Results: θ consistently identified high gradient regions in the comparison image, allowing for dynamic region of interest selection. |θ| ≤ 30° corresponded to a reference-comparison intravoxel gradient ≥ 104 HU/mm. The sensitivity limit of a θ-restricted λ was found to be object offsets ≤ 0.2 mm. The noise level was found to heavily impact λ, necessitating an optimal choice of smoothing kernel.

Conclusion: A θ-restricted λ successfully detects intravoxel differences between simulated vessels in the presence of noise. Integrating λ into image registration algorithms may improve performance at the sub-voxel level.


Cone-beam CT


IM- Cone Beam CT: Registration

Contact Email