1151 - Limited-Angle CBCT Image Reconstruction for Non-Coplanar Radiation Therapy via Missing Projection Generation Using Dataset-Free Deep Learning

Purpose/Objective(s): In non-coplanar radiation therapy (NCRT), potential collisions between the gantry and the treatment couch restrict the acquisition of CBCT to only limited-angle (LA). The incomplete angular acquisition of CBCT projections and the non-coplanar imaging geometry create challenges in reconstructing high-quality volumetric images to meet the needs of NCRT. This work aims to develop an efficient missing-projection generator that enables complete angular coverage of CBCT projections and can be seamlessly integrated into clinically available reconstruction engines to produce high-quality 3D images.

Materials/Methods: We develop an implicit neural representation (INR) network for missing projection generation. The network, based on an 8-layer MLP, takes a 3D coordinate grid as input—2D in-plane pixel locations and 1D CBCT gantry angles—and outputs projection intensity values. We pretrain the INR network with full-view (FV) projections from a patient at a reference position (couch 0^o). The training loss consists of projection-domain differences between the network’s output and the FV projections and image-domain differences from back-projected projections. In testing, LA projections from the patient's NC setup are aligned with training projections based on CBCT scan angles to update the pretrained network for generating virtually FV projections. We evaluate the performance of projection generation via comparing the quality of reconstructed images derived from the projections.

Results: A head phantom was scanned with a Varian medical linear accelerator machine. The default FV scan with 200^o coverage in half-rotation mode was used for training. The LA projections with angular coverages ranging from 15^o to 90^o acquired at a NC setup with a 10^o couch rotation were acquired for testing. We used the classic FDK method to reconstruct 3D images from both the generated FV projections and the acquired LA projections. The ground-truth image was obtained with FV acquisitions for evaluation. The PSNR and SSIM metrics in Table 1 indicate that reconstruction quality dramatically declines with reduced angle coverage in LA projections. However, the proposed generator effectively produces FV projections, enabling stable reconstruction even with extremely limited scan angle.

Conclusion: We propose a novel projection generator for LA-CBCT scan in NCRT. The reconstructed 3D images from the generated full-view projections demonstrated accuracy and stability of the proposed approach.