Seg-MambaDiff: Skin Lesion Segmentation Based on Mamba and Diffusion Model

With the advancement of medical imaging, skin lesion segmentation has become essential in clinical diagnosis and treatment, especially for early detection of dermatological conditions. Although CNNs and Transformer architectures have been widely adopted for medical image segmentation tasks, they still struggle to accurately capture subtle details. Recently, both diffusion models and Mamba models have each shown notable advancements in image segmentation tasks, however, an efficient hybrid model is still lacking, and the diffusion model continues to face the challenge of high complexity. To address these challenges, we propose Seg-MambaDiff, a novel framework for skin lesions segmenting that integrates the diffusion architecture with Mamba model. This framework combines the noise handling capabilities of the diffusion model with the efficient modeling strengths of the Mamba model. Additionally, a conditional model is incorporated for feature fusion, and the global and local feature extraction modules are introduced. Evaluations on public datasets show that Seg-MambaDiff achieves superior segmentation accuracy, significantly improving mIoU and DSC scores. Furthermore, the model achieves fewer parameters. These advancements highlight the potential of Seg-MambaDiff in addressing existing challenges and enhancing both segmentation accuracy and processing speed.