Skin disease segmentation method based on network feature aggregation module and edge enhanced attention mechanism

• Published in: IEEE access Vol. 11; p. 1
• Main Authors: Wang, Haoran, Yu, Kun, Gao, Songyuheng, Li, Qiangqiang, Guan, Qianjun
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Attention mechanism; Blood vessels; Coding; Decoding; Diseases; Encoding; Feature extraction; Feature maps; Image segmentation; Lesions; Medical imaging; Modules; Neural networks; RSUnet; Skin diseases; SwinTransformer; Task analysis; Unet
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3330379

This article proposes a skin disease segmentation network RSUnet based on network feature aggregation module and edge enhanced attention mechanism. The network subject adopts an encoding decoding structure, dividing the left encoding layer into four layers according to the order of feature map resolution from large to small. Multiple convolutional layers are used to capture the contextual information of the lesion image and extract image features. Using ResNext as the backbone feature extraction network and SwinTransformer as the auxiliary feature extraction network, fully absorbing the advantages of CNN and Transformer. In order to better combine the two and improve the segmentation of skin disease images, which may include hair, blood vessels, bubbles, etc., causing errors in the segmentation results. Propose a network feature aggregation module and input it into the next layer of feature extraction convolutional block bottleneck in ResNext. Secondly, considering that the edge shape of skin diseases is variable and fuzzy, and the segmentation task is difficult, an edge enhancement attention mechanism is proposed to increase the weight allocation of edges and enhance the edge details of the lesion area. The RSUnet designed based on the above ideas achieved better Jaccard, accuracy, recall, and dice values on the ISIC2018 dataset for skin disease segmentation compared to neural networks such as Unet, Unet++, DeepLabV3+, PSPNet, and FPN.