In this paper, we propose a novel criss-cross attention based multi-level fusion network to segment gastric intestinal metaplasia from narrow-band endoscopic images. Our network is composed of two sub-networks including criss-cross attention based feature fusion encoder and feature activation map guided multi-level decoder. The former one learns representative deep features by imposing attention on features of multiple receptive fields. The latter one segments gastric intestinal metaplasia regions by using the feature activation map scheme to enhance the importance of decoder features and avoid overfitting. As shown in the experimental results, our method outperforms state-of-the-art semantic segmentation methods on a novel challenging endoscopic image dataset.

Citation

C.-M. Nien, E.-H. Yang, W.-L. Chang, H.-C. Cheng, and C.-R. Huang, "Criss-cross attention based multi-level fusion network for gastric intestinal metaplasia segmentation," in Proc. Int. Conf. Med. Image Comput. Comput. Assist. Intervent. Workshop Imag. Syst. GI Endoscopy, 2022, pp. 13–23.

Acknowledgements

This work was supported in part by the National Science and Technology Council, Taiwan under Grant MOST 110-2634-F-006-022, 111-2327-B-006-007, and 111-2628-E-005-007-MY3. We would like to thank National Center for High-performance Computing (NCHC) for providing computational and storage resources.

Particular Thanks

Department of Internal Medicine and Institute of Clinical Medicine, National Cheng Kung University Hospital, College of Medicine, National Cheng Kung University, Tainan, Taiwan.