1230 - Dynamic CT Reconstruction Using a Latent Diffusion Model Integrating Physical Mechanisms and Population-Individual Collaborative Fine-Tuning

Purpose/Objective(s): Current models for dynamic CT reconstruction fail in clinical settings due to physical discrepancies between idealized training DRRs and real 2D kV projections. To bridge this domain shift, we propose a physics-enhanced Latent Diffusion Model (LDM) utilizing Population-Individual Collaborative Fine-Tuning to reconstruct high-fidelity dynamic CTs from single-view kV projections.

Materials/Methods: A retrospective, multi-cohort dataset was curated, comprising: (1) 1000 lung cancer patients with pre-treatment 4DCTs to provide population-level respiratory motion priors; (2) 100 nasopharyngeal carcinoma patients with paired 3DCTs at T₁ and T₂ to capture longitudinal anatomical deformation priors; and (3) 10 patients with real on-treatment 2D kV projections and perfectly paired in-room Fan-Beam CT for ground-truth clinical validation. The proposed framework operates in three stages: First, a physics-forward disentanglement module isolates anatomical features from modality-specific artifacts by simulating Compton scatter and Poisson-Gaussian noise, bridging the simulated-to-real domain gap. Second, the LDM undergoes large-scale population pre-training (lung and NPC cohorts) to learn universal 3D spatial representations and spatio-temporal motion manifolds. Finally, PSR is applied using the individual patient's limited pre-treatment CT, anchoring the model to patient-specific high-frequency topologies. Spatio-temporal mapping is achieved via 3D-Rotary Position Embedding.

Results: The mean absolute error of soft tissue regions from [45 ± 33 HU] to [19 ± 12 HU] and improving the Contrast-to-Noise Ratio by [42.5%]. Compared to conventional person-independent models, our PSR strategy effectively handled respiratory baseline shifts, reducing the tumor target centroid deviation from [3.2 ± 1.5 mm] to [0.8 ± 0.3 mm] and achieving a diaphragm motion trajectory correlation coefficient of [r = 0.912]. On the independent validation cohort of 10 patients with real paired kV and FBCT, the generated dynamic CTs achieved superior structural fidelity, with an average Structural Similarity Index (SSIM) of [0.92 ± 0.02] and a Peak Signal-to-Noise Ratio (PSNR) of [35.4 ± 1.8 dB]. Dosimetric validation confirmed high clinical safety, with Gamma passing rates (3mm/3%) consistently exceeding [96.5%] within the planning target volume (PTV) and maximum dose deviations to organs at risk restricted to under [2.0%].

Conclusion: Integrating physics-informed domain disentanglement with a PSR-driven latent diffusion framework demonstrates feasibility in mitigating the DRR-to-kV physical gap and addressing over-time patient anatomical variations. With clinically acceptable structural fidelity and dosimetric accuracy, this approach offers a potential tool for reliable intra-fractional volumetric monitoring in adaptive radiotherapy.