Skin lesion segmentation is a necessary step in the diagnosis of skin diseases.Affected by the irregular shape of the lesion area,uneven color,and fuzzy boundary,it is very difficult to accurately segment skin lesion.To address this problem,the paper proposes DSLK LK-UNet(Dilation Stage Large Kernel UNet)for skin lesion segmentation.To fuse the multi-scale information,a cavity step connection module is designed and embedded into the encoder and decoder module for effective capture the contextual information.Then,an efficient convolution module based on large kernel convolution fusion(LKF)is proposed to optimize the extrac-tion of the detail information of the small targets.Finally,an edge loss function based on Laplace is used to im-prove the detection of the skin lesion with weak boundaries.Experimental results tested on the ISIC 2018 data set show that the proposed method effectively segment skin lesions,and receives an average similarity coeffi-cient(Dice),mean intersection over union(MIoU),accuracy and F1-Score of 92.86％,89.10％,97.00％and 89.28％,respectively,which is superior to some existing segmentation methods.In addition,it also shows its good performance on the detection of the lesion region with weak boundaries.Compared with other segmenta-tion method,the proposed method has advantages in segmenting skin lesions with blurred edge and low con-trast.
万方数据
维普
