Author: Hongyu An, Phillip Cuculich, H Michael Gach, Yao Hao, Trevor McKeown, Clifford Robinson, Yuhao Wang, Deshan Yang 👨🔬
Affiliation: Washington University School of Medicine in St. Louis, Washington University School of Medicine, Duke University, Department of Radiation Oncology, Duke University, Washington University in St. Louis, Department of Radiation Oncology, Washington University School of Medicine in St. Louis 🌍
Purpose: Cardiorespiratory motion management is crucial in stereotactic arrhythmia radiotherapy to define target margins and minimize cardiac toxicity. While respiratory 4DCT (r4DCT) images can infer the respiratory of the heart motion, they suffer from phase irregularities and random cardiac motion-induced artifacts, complicating respiratory motion quantification. We developed a group-wise non-rigid iterative closest point ICP registration algorithm with an L2 smoothness penalty (L2SP) and principal component reduction (PCR) to improve phase-dimension motion smoothness, continuity, and accuracy and handle these artifacts.
Methods: L2SP and PCR were integrated into the group-wise non-rigid ICP algorithm to register target shapes to moving shapes from r4DCTs. L2SP applied L2 regularization to the deformable vector field (DVF) in the phase dimension to smooth the circular motion trajectory across the respiratory cycle. PCR used principal component analysis on the DVF to retain only two main respiratory principal components (PCs), removing less relevant PCs caused by random cardiac motion. Simulated respiratory 4DCTs from the 4D extended cardiac-torso phantom under free-breathing with simultaneous heartbeats were used to assess registration accuracy. The algorithm was tested in four conditions: without L2SP and PCR, with L2SP only, with PCR only, and with both. Target registration errors (TRE) and acceleration changes (AC) were calculated for accuracy and smoothness. Fourier transform evaluated cardiac motion reduction.
Results: TREs were 0.78 ± 0.02 mm, 0.79 ± 0.02 mm, 0.98 ± 0.07 mm, and 0.98 ± 0.08 mm for registrations without L2SP and PCR, with L2SP, with both, and with PCR, respectively. ACs were 60.7, 50.8, 27.6, and 29.8. The ground truth AC was 31.5. PCR reduced cardiac motion artifacts, while L2SP improved smoothness with minor accuracy loss.
Conclusion: The new group-wise non-rigid ICP algorithm achieved accurate image registration with TRE averaging below 1 mm. PCR reduced motion artifacts, and L2SP enhanced smoothness with slight accuracy trade-offs.