Skin Lesions Segmentation Method Based on Diffusion Model

皮肤病变分割在临床诊断中扮演着重要角色。为了加强对多尺度上下文信息的学习，提高模型在皮肤病变边缘处的分割准确度，本研究提出一种新的基于扩散模型的皮肤病变分割框架DCA-SegDiff。该框架将双交叉注意模块集成到网络的跳跃连接部分，能够同时接收来自不同层编码器的多尺度特征信息，同时与解码器部分的挤压激励机制结合使用，以充分利用上下文特征信息。同时，采用基于Haar小波的下采样模块，以保留更多边缘细节信息。此外，在编码器深层采用高效增强多尺度注意力模块，通过跨空间学习的方法来融合多尺度特征信息。为验证模型的优越性，在ISIC2018、PH2、HAM10000三个皮肤病变分割数据集上进行实验，主要评价指标Dice分别达到了0.8944、0.9446、0.9445。结果表明，DCA-SegDiff模型在三个数据集上表现优于其他现有模型，而且相比基线模型的参数量显著减少，证明了其在皮肤病变分割任务中的有效性和泛化能力。Segmentation of skin lesions plays an important role in clinical diagnosis. In this paper, a new skin lesion segmentation framework based on diffusion model, DCA-SegDiff, is proposed to enhance the learning of multi-scale context information to improve the segmentation accuracy of the model at the edge of skin lesions. The dual cross-attention module is integrated into the skip connection part of the denoising network in this framework, which can receive multi-scale feature information from different encoders at the same time. Meanwhile, the down-sampling module based on Haar wavelet is adopted to retain more edge detail information. In addition, an efficient enhanced multi-scale attention module is used in the deep layer of the encoder to integrate multi-scale feature information by cross-space learning method. In order to verify the superiority of the model, experiments are conducted on three skin lesion segmentation datasets, ISIC2018, PH2, and HAM10000, and Dice reaches 0.8944, 0.9446, and 0.9445, respectively. The results show that the DCA-SegDiff model outperforms other existing models on the three datasets, and the number of parameters is significantly reduced compared to the Baseline, demonstrating its effectiveness and generalization ability in the skin lesion segmentation task.