使用深度學習在窄帶影像和白光大腸鏡檢查中進行實時大腸息肉分割

Real-time Colorectal Polyp Segmentation with Deep Learning in NBI and WL Colonoscopy

陳玠妏(Chieh-Wen Chen)

指導教授 ： 莊曜宇

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202104216

摘要

大腸鏡檢查是一種有效預防大腸癌的技術，它可以發現和切除大腸息肉。切除大腸息肉可以有效的降低罹患大腸癌的風險。但是大腸鏡檢查的息肉漏檢率較高且受多種因素影響。因此，開發計算機輔助方法以支持內視鏡醫師準確分割結腸鏡影片中的息肉具有重要的意義。目前，深度學習的方法已應用於息肉分割。然而，息肉分割仍然是一個具有挑戰性的問題，因為息肉具有多種大小、顏色和紋理。並且息肉與其周圍粘膜之間的邊界並不明顯。為了解決這些問題，我們提出了一種改進的息肉分割模型，稱為 DarkraNet，用於精確分割大腸鏡影像中的息肉邊界。它是由編碼器-解碼器架構、反向注意模塊和後處理組成。此外，為了臨床實用性，我們將窄帶影像結腸鏡檢查影像添加到訓練數據中。對五個廣泛使用的息肉分割基準數據集的定量實驗結果表明，與現有方法相比，所提出的 DarkraNet 實現了最先進的分割精確度，並在準確性、通用性和臨床適用性方面進一步提升效用。此外，DarkraNet 是能夠實時分割息肉的。

關鍵字

深度學習 ； 息肉分割 ； 大腸鏡檢查 ； 結直腸癌 ； 窄帶影像

並列摘要

Colonoscopy is an efficient technique for the detection and removal of colorectal polyps. The removal of colorectal polyps can lower the risk for getting colorectal cancer (CRC). However, the missing rate of polyps in colonoscopic examinations is relatively high due to many factors. Thus, developing computer-aided methods to support endoscopists to accurately segment polyps in colonoscopy videos is quite crucial. Currently, deep learning approaches have been applied to polyp segmentation. However, polyp segmentation is still a tricky issue since polyps have a diversity of sizes, colors and textures. In addition, the boundary of surrounding mucosa around a polyp is not obvious. To solve these issues, we propose an improved polyp segmentation model called DarkraNet for precisely segment polyps’ boundary in colonoscopy images. It consists of an encoder-decoder architecture, reverse attention module and post-processing. Furthermore, for clinical usefulness, we add narrow-band imaging (NBI) colonoscopy images into training data. Quantitative evaluation on five widely-used polyp segmentation benchmarks datasets shows that the proposed DarkraNet achieves state-of-the-art (SOTA) segmentation accuracy compared to existing methods and further improve in terms of accuracy, generalizability and clinical applicability. Moreover, DarkraNet is able to segment polyps in real-time level.

並列關鍵字

deep learning ； polyp segmentation ； colonoscopy ； colorectal cancer ； narrow-band imaging

參考文獻

[1] Health Promotion Administration, Ministry of Health and Welfare. (2021). “Overview of Colorectal Cancer Prevention and Treatment.” https://www.hpa.gov.tw/Pages/Detail.aspx?nodeid=615 pid=1126

[2] Subramanian, V., Ragunath, K. (2014). Advanced Endoscopic Imaging: A Review of Commercially Available Technologies. Clinical Gastroenterology and Hepatology, 12(3), 368–376.e1. https://doi.org/10.1016/j.cgh.2013.06.015

[3] Machida, H., Sano, Y., Hamamoto, Y., Muto, M., Kozu, T., Tajiri, H., Yoshida, S. (2004). Narrow-Band Imaging in the Diagnosis of Colorectal Mucosal Lesions: A Pilot Study. Endoscopy, 36(12), 1094–1098. https://doi.org/10.1055/s-2004-826040

[4] Vișovan, I. I., Tanțău, M., Pascu, O., Ciobanu, L., Tanțău, A. (2017). The Role of Narrow Band Imaging in Colorectal Polyp Detection. Bosnian Journal of Basic Medical Sciences. https://doi.org/10.17305/bjbms.2017.1686

[5] Leifeld, L., Rogler, G., Stallmach, A., Schmidt, C., Zuber-Jerger, I., Hartmann, F., Plauth, M., Drabik, A., Hofstädter, F., Dienes, H. P., Kruis, W., Büttner, R., Löser, H., Drebber, U., Dignass, A., Terjung, B., Sauerbruch, T., Schreiber, S., Lanyi, B., . . . Walldorf, J. (2015). White-Light or Narrow-Band Imaging Colonoscopy in Surveillance of Ulcerative Colitis: A Prospective Multicenter Study. Clinical Gastroenterology and Hepatology, 13(10), 1776–1781.e1. https://doi.org/10.1016/j.cgh.2015.04.172