Author: Anna E. Rodrigues, Yuhao Wang, Deshan Yang 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University 🌍
Purpose: Accurate motion margin definition in StereoTactic Arrhythmia Radiotherapy (STAR) requires accurate cardiorespiratory motion assessment. However, respiratory 4DCT (r4DCT) images are affected by phase irregularities and random cardiac motion artifacts, complicating respiratory motion quantification. We developed a group-wise Coherent Point Drift shape registration algorithm gCPD-GMM-FEM to register the heart segmented in all r4DCTs together. It used Gaussian Mixture Model to model vertex position uncertainties and Finite Element Model to regularize volumetric deformations. A novel L2 smoothness penalty (L2SP) and a principal component reduction (PCR) process were implemented to enhance phase-dimension motion smoothness and continuity to account for the phase irregularities and random cardiac motion artifacts.
Methods: gCPD-GMM-FEM registers heart shapes segmented in r4DCTs. L2SP applies L2 regularization to smooth the cyclic motion trajectories of phase dimension. PCR employs principal component analysis to filter out less relevant principal components caused by cardiac motion artifacts. The algorithm was validated using simulated r4DCTs from a XCAT phantom under free breathing with simultaneous heartbeats. Registration accuracy was assessed under four settings: without L2SP/PCR, with L2SP only, with PCR only, and with both. Target registration errors (TRE) and acceleration changes (AC) were evaluated and compared.
Results: TREs were 0.32 ± 0.02 mm, 0.34 ± 0.02 mm, 0.81 ± 0.17 mm, and 0.78 ± 0.16 mm across the four conditions. AC values, which measure the phase dimension deformation regularity and smoothness, were 62.5, 52.4, 24.6, and 31.4, compared to the ground truth AC value of 31.5. Our results confirmed that PCR effectively reduced cardiac motion artifacts in r4DCTs, while L2SP improved the DVF cyclic smoothness with minimal accuracy loss.
Conclusion: The gCPD-GMM-FEM algorithm achieved sub-millimeter registration accuracy, effectively reduced random cardiac motion artifacts in the respiratory 4DCTs while enhancing the phase-dimension motion smoothness with minor accuracy trade-offs.