10 February 2026 in IEEE Transactions on Radiation and Plasma Medical Sciences
We present a novel framework for CBCT-to-MDCT translation, grounded in the Schrödinger Bridge (SB) formulation, which integrates GAN-derived priors with human-guided conditional diffusion. Unlike conventional GANs or diffusion models, our approach explicitly enforces boundary consistency between CBCT inputs and pseudo-targets, ensuring both anatomical fidelity and perceptual controllability. Binary human feedback is incorporated via classifier-free guidance (CFG), effectively steering the generative process toward clinically preferred outcomes. Through iterative refinement and tournament-based preference selection, the model internalizes human preferences without relying on a reward model. Subtraction image visualizations reveal that the proposed method selectively attenuates shading artifacts in key anatomical regions while preserving fine structural detail. Quantitative evaluations further demonstrate superior performance across RMSE, SSIM, LPIPS, and Dice metrics on clinical datasets—outperforming prior GAN- and fine-tuning-based feedback methods—while requiring only 10 sampling steps. The present validation was conducted exclusively on head and neck CBCT datasets, which represent the most clinically relevant use case for dental and craniofacial imaging. These findings underscore the effectiveness and efficiency of our framework for real-time, preference-aligned medical image translation.
1 January 2025 in IEEE Access
26 October 2026 in Physics in Medicine & Biology
Objective. Sparse-view computed tomography (CT) reduces radiation dose and acquisition time but introduces structured streak artifacts that make reconstruction severely ill-posed. Cold-diffusion approaches can model such deterministic degradation processes, but existing sampling strategies may rely on scalar similarity criteria and predefined thresholds to trigger sample-adaptive resets of the degradation-level trajectory. We aim to replace this threshold-based sampling control with spatially resolved residual conditioning. Approach. We propose ReCo-Diff, a residual-conditioned cold-diffusion framework for sparse-view CT reconstruction. At each reverse step, a zero-residual baseline prediction is re-degraded using the CT degradation operator, and the resulting signed image-domain re-degradation discrepancy is used as a spatial conditioning map for a second restoration pass. ReCo-Diff does not require an adaptive reset during residual-conditioned sampling and instead follows a predetermined degradation schedule. A one-time level transition is additionally used for highly sparse inputs. The method was evaluated on the AAPM Low-Dose CT dataset under 18-, 36-, and 72-view settings using a patient-wise split with one held-out test volume. Main results. ReCo-Diff achieved PSNR values of 38.54, 42.57, and 46.98 dB at 18, 36, and 72 views, respectively. At 18 views, matched-computation analysis showed that ReCo-Diff achieved 38.54 dB compared with 38.25 dB for a two-pass zero-residual control under the same 18-NFE budget. A spatially shifted-residual control achieved 37.65 dB, indicating that effective residual conditioning benefits from spatial correspondence with the current degraded state. Significance. ReCo-Diff replaces adaptive trajectory control with step-wise spatial residual conditioning in deterministic cold diffusion. Matched-computation analysis shows that the reconstruction benefit depends on spatially aligned residual information rather than on the additional network evaluation alone. The proposed strategy provides an alternative approach to controlling error propagation under a predetermined degradation schedule without relying on sample-adaptive reset decisions.