Medical Boundary Diffusion Model for Skin Lesion Segmentation

Skin lesion segmentation in dermoscopy images has seen recent success due to advancements in multi-scale boundary attention and feature-enhanced modules. However, existing methods that rely on end-to-end learning paradigms, which directly input images and output segmentation maps, often struggle with extremely hard boundaries, such as those found in lesions of particularly small or large sizes. This limitation arises because the receptive field and local context extraction capabilities of any finite model are inevitably limited, and the acquisition of additional expert-labeled data required for larger models is costly. Motivated by the impressive advances of diffusion models that regard image synthesis as a parameterized chain process, we introduce a novel approach that formulates skin lesion segmentation as a boundary evolution process to thoroughly investigate the boundary knowledge. Specifically, we propose the Medical Boundary Diffusion Model (MB-Diff), which starts with a randomly sampled Gaussian noise, and the boundary evolves within finite times to obtain a clear segmentation map. First, we propose an efficient multi-scale image guidance module to constrain the boundary evolution, which makes the evolution direction suit our desired lesions. Second, we propose an evolution uncertainty-based fusion strategy to refine the evolution results and yield more precise lesion boundaries. We evaluate the performance of our model on two popular skin lesion segmentation datasets and compare our model to the latest CNN and transformer models. Our results demonstrate that our model outperforms existing methods in all metrics and achieves superior performance on extremely challenging skin lesions. The proposed approach has the potential to significantly enhance the accuracy and reliability of skin lesion segmentation, providing critical information for diagnosis and treatment. All resources will be publicly available at https://github.com/jcwang123/MBDiff .