Author: Geoffrey D. Hugo, Eric Laugeman, Thomas R. Mazur, Pamela Samson, Kim A. Selting, Zhehao Zhang 👨🔬
Affiliation: University of Illinois, Washington University in St. Louis School of Medicine, WashU Medicine 🌍
Purpose: To investigate the robustness of a deep learning (DL)-based 4D-CBCT motion-compensated (MoCo) reconstruction method to out-of-distribution data.
Methods: Our developed 4D-CBCT reconstruction framework integrates two pre-trained DL models: 1) an artifact-reduction model trained on a dataset of 47 free-breathing CBCT scans acquired using conventional CBCT imaging on a Varian Edge system, designed to improve the quality of initial Feldkamp–Davis–Kress (FDK)-reconstructed 4D-CBCT images, and 2) a registration model trained on a dataset of 31 sets of 4D-CT images, aimed at achieving real-time motion model generation for rapid MoCo reconstruction.
In this study, we used two out-of-distribution thorax datasets to evaluate robustness: 1) 1-minute free-breathing scans paired with 6-second breath-hold scans acquired on a Varian Halcyon system with HyperSight imaging to assess performance on patient imaging from a different platform, and 2) a canine dataset acquired using a conventional CBCT imaging system to evaluate performance in reconstructing anatomies and breathing motions different from those in the training dataset.
Results: The pre-trained artifact-reduction model effectively mitigated streaking artifacts caused by phase binning in both datasets. However, it failed to eliminate artifacts with patterns varying from streaking artifacts, including those induced by limited-angle acquisitions and metal. The subsequent MoCo reconstruction using the registration model corrected these effects and significantly improved 4D-CBCT reconstruction quality, demonstrating the robustness of the pre-trained registration model and the overall reconstruction framework.
Conclusion: Our DL-based 4D-CBCT reconstruction method demonstrated robustness across different imaging systems and anatomies. However, the artifact-reduction model itself exhibited limited generalizability to artifact patterns beyond those caused by phase binning. This suggests that incorporating processes to address a broader range of artifact patterns is needed to further improve the robustness of DL-based 4D-CBCT methods.