融合边缘增强与多尺度注意力的皮肤病变分割  

Skin Lesion Segmentation Combining Boundary Enhancement and Multi-scale Attention

作　　者：白雪飞[1] 靳智超 王文剑[1,2] 马亚楠 BAI Xuefei;JIN Zhichao;WANG Wenjian;MA Yanan(School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education(Shanxi University),Taiyuan 030006,China)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]计算智能与中文信息处理教育部重点实验室(山西大学),太原030006

出　　处：《计算机科学》2023年第4期96-102,共7页Computer Science

基　　金：国家自然科学基金(61703252,62076154,62276161,U21A20513);山西省重点研发项目(202102150401013)。

摘　　要：皮肤病变形状、颜色、大小类型多样,给皮肤病变的准确分割带来了巨大挑战。针对这一问题,提出了一种融合边缘增强与多尺度注意力的皮肤病变分割网络(BEMA U-Net)。该网络包含一个用于提取全局特征的空间多尺度注意力模块和一个用于增强病变区域边缘特征的边缘增强模块,将两种模块添加到以编码解码结构为主干的网络(U-Net)中,能够有效抑制病变图像中背景噪声的干扰并强化病灶的边缘细节。此外,设计了混合损失函数,结合Dice Loss和Boundary Loss,并在训练过程中实现混合损失函数的动态权重调整,使网络对病变图像整体特征和边缘细节特征的提取进行多重监督,缓解了皮肤病变图像分割中毛发干扰和边缘模糊的问题。在ISIC2017和ISIC2018两个公开数据集上的实验结果表明,所提算法对皮肤病变部位的分割图像边缘连续、轮廓清晰,具有更好的分割效果。In view of the various types of skin lesions in shape,color and size,which pose a huge challenge to the accurate segmentation of skin lesions,a skin lesion segmentation network that combines boundary enhancement and multi-scale attention is proposed(BEMA U-Net).It consists of two modules,one is called spatial multi-scale attention module,which is used to extract spatial global features,and the other is called boundary enhancement module,which is used to enhance the edge features of the lesion area.BEMA U-Net adds the two modules to the U-Net network with encoding and decoding structure,which can effectively suppress the interference of background noise in the image of lesions and enhance the edge details of lesions.In addition,the mixed loss function is designed,Dice loss and Boundary loss are combined,and the dynamic weight adjustment of the mixed loss function is realized in the training process,so that the network could carry out multiple supervision on the extraction of the overall features and edge details of the pathological images,and the problems of hair interference and edge blur in the segmentation of skin pathological images are alleviated.Experimental results on ISIC2017 and ISIC2018 public data sets show that the proposed algorithm has better segmentation effect for skin lesions with continuous edges and clear contours.

关 键 词：皮肤病变分割 空间多尺度注意力 全局特征 边缘增强 U-Net 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]