VS-DDPM: Efficient Low-Cost Diffusion Model for Medical Modality Translation
Researchers have introduced VS-DDPM, a novel 3D Variable-Step Denoising Diffusion Probabilistic Model designed to address the slow inference speeds typical of diffusion models while maintaining high generative quality. This framework was evaluated within the BraTS2025 and SynthRAD2025 challenges, focusing on four specific medical imaging tasks: missing MRI synthesis, tumor removal, MRI-to-sCT translation, and CBCT-to-sCT translation. The model achieved state-of-the-art performance in missing MRI synthesis, recording Dice scores of 0.80, 0.83, and 0.88 for enhancing tumor, tumor core, and whole tumor regions, respectively, with a structural similarity index of 0.95. In tumor removal tasks, it demonstrated strong metrics including an RMSE of 0.053 and PSNR of 26.77. Although competitive in cross-modality translations, it did not reach top benchmarks there, likely due to preprocessing sensitivities. The study highlights VS-DDPM as a robust, efficient solution for high-fidelity 3D medical image synthesis under hardware constraints, with open-source code provided for further research and application in clinical settings.
Wire timeline
VS-DDPM: Efficient Low-Cost Diffusion Model for Medical Modality Translation
Researchers have introduced VS-DDPM, a novel 3D Variable-Step Denoising Diffusion Probabilistic Model designed to address the slow inference speeds typical of diffusion models while maintaining high generative quality. This framework was evaluated within the BraTS2025 and SynthRAD2025 challenges, focusing on four specific medical imaging tasks: missing MRI synthesis, tumor removal, MRI-to-sCT translation, and CBCT-to-sCT translation. The model achieved state-of-the-art performance in missing MRI synthesis, recording Dice scores of 0.80, 0.83, and 0.88 for enhancing tumor, tumor core, and whole tumor regions, respectively, with a structural similarity index of 0.95. In tumor removal tasks, it demonstrated strong metrics including an RMSE of 0.053 and PSNR of 26.77. Although competitive in cross-modality translations, it did not reach top benchmarks there, likely due to preprocessing sensitivities. The study highlights VS-DDPM as a robust, efficient solution for high-fidelity 3D medical image synthesis under hardware constraints, with open-source code provided for further research and application in clinical settings.
cs.AI updates on arXiv.org