- The paper introduces VS-DDPM, a variable-step diffusion model that adjusts denoising steps based on computational constraints to achieve efficient, high-fidelity medical modality translation.
- Experimental results on MRI-to-sCT and CBCT-to-sCT tasks demonstrate competitive performance with MAEs of 72.5–73.9 and Dice indices up to 0.785.
- The study also achieves state-of-the-art results in missing modality synthesis and tumor inpainting, highlighting practical trade-offs like patch-based training and mixed-precision inference.
VS-DDPM: Efficient Low-Cost Diffusion Model for Medical Modality Translation
Introduction
VS-DDPM addresses a persistent bottleneck in medical image synthesis: how to achieve high-fidelity, reliable cross-modality translation with efficient compute and minimal inference delay. Targeted at clinical radiotherapy workflows, the method directly confronts the hardware, memory, and timing constraints prevalent in both research environments and clinical translation. The framework is evaluated across diverse MRI- and CBCT-based translation tasks, including MRI-to-sCT, CBCT-to-sCT, missing MRI modality synthesis, and tumor inpainting, reflecting core needs in radiotherapy planning and multimodal neuroimaging.
Technical Contributions
The principal technical innovation is the Variable-Step Denoising Diffusion Probabilistic Model (VS-DDPM), extending DDPMs (2604.22942) by introducing support for variable sampling trajectories (T) during inference. This enhancement enables the model to adapt the number of denoising steps in accordance with computational budgets and input dimensionality, without retraining. This contrasts with both canonical DDPMs, which typically use fixed T, and DDIM, which empirically underperformed in this regime.
Model architectures are based on 3D UNet and 3D Swin-ViT backbones, with training employing patch-based strategies and sliding window inference to accommodate hardware constraints. Careful attention is paid to data normalization and augmentation strategies tailored to the modality and anatomical region. The loss function is a composite of MAE, MSE, SSIM, and VLB regularization; in the sCT translation tasks, AFP (Anatomical Feature-Prioritized) loss is additionally employed for fine-tuning, reflecting the prioritization of structure fidelity over per-voxel similarity.
A pragmatic procedural detail is the dynamic adjustment of T at inference, derived from real-time latency measurements and scan dimensions to respect strict real-world inference budgets (e.g., under 15 minutes per volume on an NVIDIA T4 GPU).
Experimental Results
MRI/CBCT to sCT Translation
VS-DDPM demonstrates competitive performance on the SynthRAD2025 challenge datasets for both MRI-to-sCT and CBCT-to-sCT translation, achieving MAEs of 72.5–73.9 (MRI-to-sCT, variable-step, U-Net) and 58.3–60.9 (CBCT-to-sCT, variable-step, U-Net), respectively. MS-SSIM and Dice indices further corroborate strong structural fidelity (MS-SSIM ≈ 0.915, Dice ≈ 0.71 for MRI; MS-SSIM ≈ 0.953, Dice ≈ 0.785 for CBCT). These numbers are on par with, but do not surpass, ensemble-based SOTA solutions benchmarked on the SynthRAD2025 platform, which report lower MAE and higher similarity indices.
The model reveals weak sensitivity to the number of denoising steps when using variable-step inference, supporting the claim that VS-DDPM is robust against adaptive inference schedules. However, the authors highlight the practical trade-offs: patch-based training and mixed-precision inference, critical for efficiency, may contribute to a slight metric deficit and training instability relative to full-volume, high-precision approaches.
Missing Modality Generation
In the BraTS2025 missing MRI modality challenge, VS-DDPM achieves SOTA results, with Dice scores of 0.80/0.83/0.88 (enhancing tumor/core/whole tumor) and SSIM of 0.95. The system preserves >95% of the segmentation performance relative to real multimodal input, with only marginal reduction in downstream segmentation DSC. The pipeline also exhibits minimal blurring or information loss, as reflected by high SSIM, especially after post-processing to suppress low-intensity background artifacts.
Tumor Removal (Inpainting)
On BraTS-based tumor removal, online test set evaluation establishes that VS-DDPM attains highly competitive (joint SOTA) results—RMSE of 0.053, PSNR of 26.77, SSIM of 0.918. The reconstructions are visually sharper than those from competing U-Net baselines, with greater retention of high-frequency detail and reduced oversmoothing.
Implications for Practice and Theory
VS-DDPM's capacity to maintain SOTA or near-SOTA performance while tuning inference and resource demand downstream is significant for real-world clinical adoption. The framework provides a viable path for integration of diffusion models into routine workflows on cost-constrained or time-constrained hardware. The ability to synthesize fully-populated MRI protocols, generate radiation-free sCTs, and perform lesion inpainting with high anatomical fidelity is directly relevant to adaptive treatment planning, diagnostic support, and augmentation of downstream neuroimage processing algorithms.
The work also underscores critical engineering trade-offs inherent in 3D diffusion modeling for volumetric medical data: the window-based approach and loss landscape may limit global context and introduce overfitting risks. The training instability with mixed-precision further points to the need for robust, resource-aware optimization in the medical domain.
Future Directions
Further development is required for enhanced training stability, reduction of patch-based inference artifacts, and maximization of fidelity under extreme compute constraints. Full-volume end-to-end architectures, subject to hardware advances, could obviate current windowing compromises. The model’s variable-step schedule mechanism may also generalize to other generative modeling applications within and beyond medical imaging.
Exploration of improved data harmonization and pre/post-processing protocols—as flagged by the sCT translation performance gap—remains necessary. The biophysical validity of synthesized modalities and their impact on clinical downstream tasks warrants continued longitudinal and multi-institutional evaluation.
Conclusion
VS-DDPM offers an efficient, hardware-adaptive diffusion model architecture for medical modality translation, achieving robust and high-performing results on missing modality imputation and tumor inpainting, and strong competitive performance on MRI/CBCT-to-sCT translation. The framework’s flexibility in step scheduling and compute adaptation positions it as a practical candidate for clinical and research deployment, with future work needed to further optimize stability, fidelity, and extensibility.
