KEY POINTS
- This retrospective technical study included 360 cone-beam computed tomography datasets from 15 patients receiving image-guided gynecologic radiotherapy. For each patient, 20 fractions were used for patient-specific training and 4 fractions for testing, yielding 60 held-out test volumes.
- The physics-constrained dual-domain network reconstructed volumetric cone-beam computed tomography from simulated anterior–posterior and lateral radiographs. It combined projection restoration, analytic cone-beam backprojection, and volumetric diffusion-based refinement.
- The proposed model achieved a mean peak signal-to-noise ratio of 48.02 dB, structural similarity index of 0.9880, and mean absolute error of 24.89 HU.
- Performance exceeded both the conditional generative adversarial network (37.94 dB, 0.9546, 48.45 HU) and unconstrained diffusion model (38.89 dB, 0.9634, 46.29 HU); all comparisons with the proposed model were significant at p<0.001.
- Testing used digitally reconstructed projections derived from existing cone-beam computed tomography rather than measured onboard radiographs. The approach also required approximately 20 prior fractions for patient-specific training and was not validated for dose recalculation, adaptive replanning, or margin reduction.
CLINICAL TAKEAWAY
The framework could eventually provide a volumetric anatomical surrogate when full-arc cone-beam computed tomography is impractical, while retaining the speed and lower imaging burden of orthogonal radiographs. However, it requires validation with real clinical projections, earlier-course deployment, and dose-critical tasks before clinical use; this is technically relevant, not practice-changing.