Attention-Based DenseUnet Network With Adversarial Training for Skin Lesion Segmentation

• Published in: IEEE access Vol. 7; pp. 136616 - 136629
• Main Authors: Wei, Zenghui, Song, Hong, Chen, Lei, Li, Qiang, Han, Guanghui
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: adversarial training; Attention; Datasets; DenseUnet; Feature extraction; Feature maps; Image segmentation; Information transfer; Lesions; Matching; Melanoma; Modules; Shape; Skin; skin lesion segmentation; Training
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2019.2940794

Automatic segmentation of skin lesions in dermoscopy images is a challenging task due to the large size and shape variations of the lesions, the existence of various artifacts, the low contrast between the lesion and the surrounding skin. In this paper, we propose a novel Attention Based DenseUnet network (referred as Att-DenseUnet) with adversarial training for skin lesion segmentation. Att-DenseUnet is a Generative Adversarial Network which contains two major components: Segmentor and Discriminator. In the Segmentor module, we propose an architecture which is similar to DenseNet in the down-sampling path to ensure maximum multi-scale skin lesions information transfer between layers in the network at dense scale range, meanwhile, we design an attention module to automatically focus on the skin lesion features and suppress the irrelevant artifacts features in the output feature maps of the DenseBlocks. In the Discriminator module, we employ adversarial feature matching loss to train the Segmentor stably, force the Segmentor to extract multi-scale discriminative features, and guide the attention module focusing on the multi-scale skin lesions. A novel loss function of the Segmentor is proposed which combines the jaccard distance loss with the adversarial feature matching loss introduced by the Discriminator. We trained the proposed Att-DenseUnet on ISBI2017 dataset. The test results show that our approach gains the state-of-the-art performance, especially for JAC (0.8045) and SEN (0.8734) scores which are significantly improved by 2.2% and 1.9%, respectively, also our network is robust to different datasets, and gains the lowest time cost which make our network suitable for clinical application.