TANet: Triple Attention Network for medical image segmentation. (April 2023)

• Record Type: Journal Article
• Authors: Wei, Xin; Ye, Fanghua; Wan, Huan; Xu, Jianfeng; Min, Weidong
• Is Part Of: Biomedical signal processing and control. Volume 82(2023)
• Publication Date: 2023-04
• Subjects: 41A05 -- 41A10 -- 65D05 -- 65D17; Attention mechanism -- Deep learning -- Medical image processing -- Image segmentation; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104608
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: In recent years, deep learning-based methods have achieved remarkable progress in medical image processing, like polyp segmentation in colonoscopy images and skin lesion segmentation in dermoscopy images. However, the current state-of-the-art medical segmentation methods still suffer from the problem of low accuracy in segmenting the small-scale and variable-scale objects. To solve this problem, we propose Triple Attention Network (TANet). In TANet, a novel Triple Attention Module (TAM) is presented. TAM has two sub-modules: Multi-scale Feature Selection Module (MFSM) and Contextual Feature Extraction Module (CFEM). MFSM is used to extract more adaptable multi-scale features for capturing variable-scale objects, while CFEM is for capturing small-scale objects by extracting contextual features. TAM aims to combine MFSM and CFEM to finally enhance the segmentation performance of the medical images with the small-scale and variable-scale lesions. Extensive experiments are conducted on five polyp datasets and one skin lesion dataset. Results show that the proposed models outperform the previous state-of-the-art models on most evaluation metrics and improve the Dice score by up to 7.1%. All results consistently confirm the effectiveness of the proposed TANet and show that the TANet achieves state-of-the-art performance on the above datasets. Highlights: Help clinicians make the diagnosis by automatically marking the diseased tissues. Low-level features contribute limitedly but incur substantial computational cost. Correlation between channels and spaces can be used for medical image segmentation. Attention mechanism allows the networks to pay attention to the areas of Interest. Selecting appropriate scale features can solve the scale adaptability problem.