BLA-Net:Boundary learning assisted network for skin lesion segmentation. (November 2022)

• Record Type: Journal Article
• Authors: Feng, Ruiqi; Zhuo, Li; Li, Xiaoguang; Yin, Hongxia; Wang, Zhenchang
• Is Part Of: Computer methods and programs in biomedicine. Volume 226(2022)
• Publication Date: 2022-11
• Subjects: Skin lesion segmentation -- Global context information extraction module -- Auxiliary boundary learning network -- Dynamic deformable convolution; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2022.107190
• Languages: English
• ISSNs: 0169-2607

Abstract

Abstract: Background and Objective: : Automatic skin lesion segmentation plays an important role in computer-aided diagnosis of skin diseases. However, current segmentation networks cannot accurately detect the boundaries of the skin lesion areas. Methods: : In this paper, a boundary learning assisted network for skin lesion segmentation is proposed, namely BLA-Net, which adopts ResNet34 as backbone network under an encoder-decoder framework. The overall architecture is divided into two key components: Primary Segmentation Network (PSNet) and Auxiliary Boundary Learning Network (ABLNet). PSNet is to locate the skin lesion areas. Dynamic Deformable Convolution is introduced into the lower layer of the encoder, so that the network can effectively deal with complex skin lesion objects. And a Global Context Information Extraction Module is proposed and embedded into the high layer of the encoder to capture multi-receptive field and multi-scale global context features. ABLNet is to finely detect the boundaries of skin lesion area based on the low-level features of the encoder, in which an object regional attention mechanism is proposed to enhance the features of lesion object area and suppress those of irrelevant regions. ABLNet can assist the PSNet to realize accurate skin lesion segmentation. Results: : We verified the segmentation performance of the proposed method on the two public dermoscopy datasets, namely ISBI 2016 and ISIC 2018. The experimental results show that our proposed method can achieve the Jaccard Index of 86.6%, 84.8% and the Dice Coefficient of 92.4%, 91.2% on ISBI 2016 and ISIC 2018 datasets, respectively. Conclusions: : Compared with existing methods, the proposed method can achieve the state-of-the-arts segmentation accuracy with less model parameters, which can assist dermatologists in clinical diagnosis and treatment.