ABSTRACT Medical image segmentation is pivotal for precisely identifying and delineating regions of interest, as it supplies indispensable information for diagnosing and managing a wide spectrum of diseases. In recent years, diffusion probabilistic models have rapidly gained traction, with representative works such as MedSegDiff and Cold SegDiffusion surpassing mainstream MAXFormer and TransUNet architectures in segmentation accuracy and demonstrating substantial clinical utility. Nevertheless, existing DPM‐based approaches still face two prominent challenges: the degradation of accuracy caused by fixed diffusion steps and the excessive reliance on a constant proportion of the original image as conditional input. To address these limitations, we propose DyScoreDiff, a novel diffusion‐based framework that leverages dataset‐derived quality scores and integrates two core modules which dynamically adjust the number of diffusion steps and adaptively tune the proportion of conditional features, respectively. Validated on three distinct modality‐specific segmentation datasets, DyScoreDiff consistently outperforms U‐Net, Swin‐Unet, and other state‐of‐the‐art diffusion‐based methods across all evaluation metrics, underscoring its robust efficacy and strong generalizability in diverse clinical scenarios.
Hu et al. (Thu,) studied this question.