Physics-constrained and geometry-aware deep reconstruction for limited-angle CBCT

Abstract

Objective. Limited-angle cone-beam computed tomography (CBCT) is attractive for image-guided radiotherapy and imaging-guided procedures when full-arc acquisition is impractical, but the incomplete angular sampling leads to severe streaking, shading, and structural distortion. This study aimed to develop a physics-constrained and geometry-aware deep reconstruction framework for recovering anatomically faithful volumetric CBCT from short-arc limited-angle projections. Approach. We propose a physics-constrained and geometry-aware network (PCGA-Net), a dual-domain reconstruction framework that combines projection completion, analytic geometric reconstruction, and image-domain refinement within a unified cross-domain design. Specifically, a projection completion module first estimates missing-angle information from the measured limited-angle projections. A geometry-aware reconstruction module then maps the completed projections into image space using differentiable cone-beam forward and backward-projection operators. Finally, a volumetric refinement module suppresses residual artifacts and restores anatomical detail under geometry-guided cross-domain consistency supervision. The framework was evaluated on three clinically distinct cohorts: gynecologic (GYN) high-dose-rate (HDR) brachytherapy, abdominal liver four-dimensional computed tomography (4DCT), and head-and-neck (H&N) CBCT. Main results. Across all three cohorts, fully adapted PCGA-Net consistently outperformed conventional limited-angle Feldkamp–Davis–Kress (FDK) reconstruction and the compared learning-based baselines, a conditional generative adversarial network (cGAN) and PatRecon-Net, in both qualitative and quantitative evaluation under matched patient-specific adaptation conditions. On the GYN HDR brachytherapy cohort, PCGA-Net achieved a mean absolute error (MAE) of 34.47 ± 6.28 Hounsfield units (HU), a structural similarity index measure (SSIM) of 0.87 ± 0.04, and a peak signal-to-noise ratio (PSNR) of 35.79 ± 0.95 dB, compared with 178.63 ± 31.58 HU, 0.51 ± 0.03, and 18.97 ± 1.31 dB for FDK. On the abdominal liver 4DCT cohort, PCGA-Net achieved 39.27 ± 6.10 HU MAE, 0.82 ± 0.03 SSIM, and 36.85 ± 1.16 dB PSNR, versus 195.21 ± 40.08 HU, 0.50 ± 0.06, and 18.50 ± 1.50 dB for FDK. On the H&N CBCT cohort, PCGA-Net achieved 28.42 ± 4.88 HU MAE, 0.91 ± 0.05 SSIM, and 38.52 ± 0.95 dB PSNR, compared with 165.30 ± 25.40 HU, 0.61 ± 0.05, and 21.13 ± 1.14 dB for FDK. In all three cohorts, PCGA-Net also exceeded cGAN and PatRecon-Net across all evaluated metrics. Significance. These results indicate that jointly enforcing projection-domain completion, analytic imaging geometry, and image-domain refinement is an effective strategy for limited-angle CBCT reconstruction. PCGA-Net provides a robust short-arc reconstruction framework across multiple anatomical sites and clinical workflows, supporting its potential utility in constrained image-guided therapeutic and procedural settings.