SEACU-Net: Attentive ConvLSTM U-Net with squeeze-and-excitation layer for skin lesion segmentation. (October 2022)

• Record Type: Journal Article
• Authors: Jiang, Xiaoliang; Jiang, Jinyun; Wang, Ban; Yu, Jianping; Wang, Jun
• Is Part Of: Computer methods and programs in biomedicine. Volume 225(2022)
• Publication Date: 2022-10
• Subjects: Skin lesion segmentation -- U-Net -- Squeeze-and-excitation -- ConvLSTM -- Loss function; 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.2022.107076
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: An extended version of U-Net network for skin lesion image segmentation. In the last layer of the contraction path, the idea of dense convolution is introduced to facilitate the propagation of information within network. At each encoding and decoding stage, a channel and spatial squeeze & excitation layer is constructed after each convolution. An attentive ConvLSTM block is employed to integrate contextual information, which improves sensitivity and accuracy of prediction. Abstract: Background and Objective: Accurate segmentation of skin lesions is a pivotal step in dermoscopy image classification, which provides a powerful means for dermatologists to diagnose skin diseases. However, due to blurred boundaries, low contrast between the lesion and its surrounding skin, and changes in color and shape, most existing segmentation methods still face great challenges in obtaining receptive fields and extracting image feature information. To settle the above issues, we construct a new framework, named SEACU-Net, to analyze and segment skin lesion images. Methods: Inspired by the U-Net, we utilize dense convolution blocks to obtain more discriminative information. Then, at each encoding and decoding stage, a channel and spatial squeeze & excitation layer are designed after each convolution, to adaptively enhance useful information features and suppress low-value ones from different feature channels. In addition, the attention mechanism is integrated into the convolutional long short-term memory (ConvLSTM) structure, which improves sensitivity and prediction accuracy. Furthermore, this network introduces a novel loss based on binary cross-entropy and Jaccard losses, which can ensure more balanced segmentation. Results: The proposed method is applied to the ISIC 2017 and 2018 publicly image databases, then obtains a better performance in Dice, Jaccard, and Accuracy, with 89.11% and 87.58% Dice value, 80.50% and 78.12% Jaccard value, 95.01%, and 93.60% Accuracy value, respectively. Conclusion: The results of quantitative and qualitative experiments show that our method reaches high-performance skin lesion segmentation, and can help radiologists make radiotherapy treatment plans in clinical practice.