Feasibility of cone beam computed tomography with invertible recurrent inference machines
MetadataShow full item record
Introduction: Online adaptive radiation therapy (ART) relies on high-quality in-room volumetric imaging. Conventional linear accelerators (linac) are generally equipped with onboard orthovoltage cone beam computed tomography (CBCT) systems. Due to their suboptimal image quality, however, their use is limited to patient positioning tasks. Deep learning-based methods can potentially generate synthetic-CTs (sCT) directly from CBCT projection data, and enable CBCT-based online ART to be brought to conventional treatment delivery systems. Designed for solving inverse problems, invertible recurrent inference machines (iRIM) can be considered a suitable candidate for this task. To this end, the feasibility of using iRIM frameworks for CBCT image reconstruction is examined in this work. Materials & Methods: 2D and 3D iRIM models were implemented and trained on datasets composed of CBCTs from sixty-two head and neck patients who underwent head and neck image-guided radiation therapy (IGRT). In the 2D models, the CBCT inverse problem was approximated as a parallel-beam CT, while in their 3D counterparts, a CBCT geometry was emulated. In parallel, trainings using sparse CBCT data were performed on both 2D and 3D models by applying two-fold and four-fold reductions in acquisition angles. Utilizing the structural similarity index measure (SSIM) as a metric, the performances of the trained models were evaluated and subsequently compared using the paired Mann-Whitney U test. Results: The 2D and 3D iRIMs respectively achieved SSIMs (mean ± std) of (0.96 ± 0.02) and (0.94 ± 0.04) in the case of the complete CBCT acquisitions. The two-fold and four-fold reduction in acquisition angles yielded (0.94 ± 0.04) and (0.93 ± 0.05) for the 2D iRIM, and (0.93 ± 0.05) and (0.92 ± 0.05) for the 3D iRIM, respectively. In all cases, the performances of the 2D iRIMs were superior to their corresponding 3D ones, with the differences not being found significant. Conclusion: iRIMs can be used for CBCT reconstruction in 2D and 3D cases, even for undersampled acquisitions. This makes iRIM an excellent candidate to obtain high-quality CT-grade reconstructions from CBCT data, and potentially bring online ART to conventional clinical linacs.