Abstract

Info

Author(s):: YANG Guoliang; ZOU Junfeng; LI Shicong; WEN Junlin; School of Electrical Engineering and Automation, Jiangxi University of Science and Technology, Ganzhou 341000, China

Keywords:: Keywords: melanoma image segmentation skin lesion multiscale fusion deep learning

Abstract:: Abstract: The image segmentation of skin lesions can be used as an important basis for the auxiliary diagnosis of related diseases. Considering the complex structure and uneven scale information of skin lesions, a novel skin lesion segmentation method based on U-shaped dense feature fusion is proposed. The dense network structure and atrous spatial pyramid pooling are adopted in encoder for feature extraction and fusion. The dense spatial attention module and the depthwise separable convolution are used to decode deep features to prevent noise interference around the focal area. Moreover, the segmentation performance is further improved by blend squeeze attention module, and the proposed algorithm is optimized by loss function combining binary cross entropy and Jaccard coefficient. The Jaccard similarity and Dice coefficient in the simulation evaluation on ISBI 2016 skin lesions datasets were 86.87% and 92.98%, respectively. The proposed method is conducive to improving the diagnosis efficiency of skin lesions.