A comparative study of attention mechanism based deep learning methods for bladder tumor segmentation. (March 2023)

• Record Type: Journal Article
• Authors: Zhang, Qi; Liang, Yinglu; Zhang, Yi; Tao, Zihao; Li, Rui; Bi, Hai
• Is Part Of: International journal of medical informatics. Volume 171(2023)
• Publication Date: 2023-03
• Subjects: Deep learning -- Image segmentation -- Intelligent diagnosis -- Bladder cancer -- Attention mechanism; Medical informatics -- Periodicals; Information science -- Periodicals; Computers -- Periodicals; Medical technology -- Periodicals; Medical Informatics -- Periodicals; Technology, Medical -- Periodicals; Computers; Information science; Medical informatics; Medical technology; Electronic journals; Periodicals; Electronic journals; 610.285
• DOI: 10.1016/j.ijmedinf.2023.104984
• Languages: English
• ISSNs: 1386-5056

Abstract

Abstract: Background: Artificial intelligence aided tumor segmentation has been applied in various medical scenarios and showed effectiveness in helping physicians observe the potential malignant tissues. However, little research has been conducted for the cystoscopic image segmentation problem. Methods: This paper provided a comprehensive comparison of various attention modules for improving the bladder tumor segmentation performance by utilizing the cystoscopic images from Peking University Third Hospital within 2017-2022. Furthermore, this paper presented an attention mechanism based cystoscopic images segmentation (ACS) model, which was featured by the following points: (1) A mixed attention module including both the channel and spatial attention modules was integrated in the encoder-decoder path, which helped to exploit the global information of the tumor area more effectively. (2) A guidance and fusion attention module was introduced in the skip connection part, facilitating the integration of the high-level semantic features with low-level fine-grained features and the discarding of irrelevant features. (3) An inception attention module was added to enhance the feature expression in the scale of pixel level, so as to better discriminate multi-scale targets. Results: The proposed ACS model showed obviously better tumor segmentation performance than the compared models, with Dice of 82.7% and MIoU of 69% achieved. Conclusions: The proposed ACS model achieved significantly better diagnostic performance than the previous bladder tumor segmentation method based on U-Net. Our ACS model is expected to be a useful support tool to assist the tumor segmentation under cystoscopy. Highlights: The ACS model proposed in this study for cystoscopic image segmentation improves the Dice result to 82.7%. This study provides a comprehensive comparison of various attention modules for bladder tumor segmentation. The proposed ACS model can be extended for the other tumor segmentation tasks.