Purpose: To develop a joint deformable registration and reconstruction network (Reg-ReconNet) to reconstruct high-quality, field-of-view (FOV)-extended CBCT images from sparse and limited-size projections.
Methods: Reg-ReconNet leverages prior CT/CBCT information to reduce the number of cone-beam projections needed for on-board CBCT reconstruction, and to extend the effective FOV. It starts with an unsupervised 2D-to-3D deformable registration network (2D3D-RegNet), which deforms prior images to estimate on-board CBCT images by matching digitally-reconstructed-radiographs (DRRs) of the deformed prior images to acquired on-board cone-beam projections. The 2D3D-RegNet was followed with a deep cascaded CBCT reconstruction network, which features recurrent inputs of cone-beam projections at each cascade level, repeated data fidelity enforcement layers composed of embedded forward/backward projection operators, and task-specific convolutions and residual connections to correct aliasing, streaking and truncation artifacts from reconstruction. We used 14 patient cases from a TCIA lung library to train and evaluate Reg-ReconNet, each of which has multiple 4D-CT image sets acquired throughout the radiotherapy treatments. For each patient, we extracted one phase image from the first 4D-CT set as prior information, and simulated cone-beam projections from all intra-fractional/inter-fractional 4D-CTs for training/testing (10/4 patients, corresponding to 400/200 image volumes). The reconstructed images were compared against the ‘ground-truth’ original 4D-CTs using the relative-error (RE) metric.
Results: Using 5 projections, the average(±s.d.) REs of reconstructed CBCTs by the Feldkamp-Davis-Kress (FDK) algorithm, iterative algebraic-reconstruction-technique with total-variation regularization (ART-TV), 2D3D-RegNet (without following ReconNet), ReconNet (without preceding 2D3D-RegNet), and Reg-ReconNet were 198.4±37.8%, 44.2±3.8%, 20.3±5.1%, 17.8±6.8%, and 16.9±3.8%, respectively. Corresponding results were 112.8±21.8%, 27.6±1.8%, 18.1±4.8%, 11.5±4.3%, and 10.4±2.6% for 20 projections, and 78.9±14.3%, 19.7±3.2%, 16.8±4.4%, 9.1±3.2% and 7.6±1.8% for 60 projections.
Conclusion: Reg-ReconNet incorporates prior information into new CBCT reconstruction through joint deformable registration and image reconstruction. It can reconstruct high-quality images from sparse projections and extend the effective FOV by merging and updating prior information.
Funding Support, Disclosures, and Conflict of Interest: The study was supported by funding from the National Institutes of Health (R01CA240808) and from the University of Texas Southwestern Medical Center.