Abstract: Diffusion models can deliver high-quality medical image segmentation via iterative denoising, but multi-step inference often incurs prohibitive latency. We propose FMDiffBTS, a lightweight ...