Purpose: 4D-CBCT is valuable for localizing targets affected by respiratory motion in radiotherapy. However, it is limited by long scanning time, high imaging dose, and degraded image quality due to the under-sampling in each respiratory phase. Compressed sensing-based methods suffer from severe blurriness, while motion-compensated methods are limited by their slow speed and compensation errors. In this study, we proposed a novel feature-compensated deformable convolutional network (FeaCo-DCN) to realize the fast 4D-CBCT with high image quality.
Methods: The proposed FeaCo-DCN consists of (1) encoding networks which extract features from each phase of 4D-CBCT reconstructed by the FDK method, (2) deformable convolutional networks which align features of other phases to those of the target phase, and (3) a decoding network which combines and decodes features from all phases to yield high-quality images of the target phase. Different phases of 4D-CBCT are reconstructed in parallel using FeaCo-DCN. We trained and tested FeaCo-DCN on 9 fast (1-minute) 4D-CBCT scans in the AAPM SPARE Challenge using the leave-one-out strategy. Results were compared to the top-ranking methods in the Challenge in both image quality and tumor localization accuracy.
Results: FeaCo-DCN generated high-quality 4D-CBCT while reducing about 90% of the conventional scanning time. Images reconstructed by FeaCo-DCN showed accurate and clear structures and achieved 3D tumor localization accuracy within 2.5 mm, considerably outperforming the compressed sensing-based and motion-compensated algorithms. Additionally, FeaCo-DCN took only about 2.5 seconds to reconstruct images of one respiratory phase.
Conclusion: FeaCo-DCN showed effectiveness in realizing fast and high-quality 4D-CBCT while reducing about 90% of the acquisition time. The proposed feature-compensated reconstruction can be applicable for other 4D imaging modalities beyond 4D-CBCT, such as 4D-MRI and 4D-PET. FeaCo-DCN can become a valuable tool for 4D imaging for motion management in radiotherapy.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Institutes of Health under Grant No. R01-CA184173 and R01-EB028324.