CTDS-Net:CNN-Transformer Fusion Network for Dermoscopic Image Segmentation
作者
Yun Jiang,Meiqi Wang,Zequn Zhang,Hao Qiao,Wei Ming Yan,Jie Chen
标识
DOI:10.1109/icrcv59470.2023.10329022
摘要
To achieve accurate skin diagnosis and therapy, automated techniques that ensure precise segmentation of skin lesion region are crucial. Challenges related to it have been effectively overcome by convolutional neural network-based segmentation techniques. However, convolutional internal operations have their limitations, which result in poor performance of pure convolutional models when dealing with long-distance dependencies. Meanwhile, the dermoscopic images and skin segmentation target vary in size and position while also potentially presence of hair in the skin images. To solve these issues, we propose a dual encoder-single decoder model that combines the Transformer architecture with a convolutional neural network (CTDS-Net). Our method employs a fusion attention module during the down-sample stage to integrate features both branches, thereby alleviating the constraints of internal convolutional operation. Furthermore, we introduce cross-skip connection to enhance the integration of shallow-deep features, allowing the up-sample process to leverage information from down-sample more effectively. Lastly, we employ a multi-scale fusion module to capture information about multi-scale features. Comprehensive experiments substantiated the CTDS-Net's superiority and evaluated on publicly available datasets: ISIC-2017, ISIC-2018 and PH2. Respectively, the Dice coefficients are 89.70%, 93.66%, and 95.59%. The ability of our model to generalize has also been tested on the segmentation polyps tasks.