使用前處理與後處理方法改善基於深度學習之電腦輔助盲腸偵測準確率

Pre- and Post- Processing Algorithms with Deep Learning Classifier for Cecum Recognition

劉雅雰(Ya-Fen Liu)

指導教授 ： 陳中平

共同指導教授 ： 邱瀚模 ； 王偉仲(Wei-Chung Wang)

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(2020年)

https://doi.org/10.6342/NTU202004233

摘要

大腸鏡檢查是近年來使用最普遍、也是診斷大腸癌的利器，品質不佳的大腸鏡檢查不僅可能錯失診斷癌前病變的先機，更可能增加不適及併發症，故大腸鏡檢查品質的重要性可見一斑，而其評估指標包含：盲腸到達率、腺瘤偵測率、腸道準備以及退出時間，鑑於台灣目前的情況，結腸鏡檢查是否達到盲腸仍以醫師申報為主，缺乏客觀的評估方法，計算盲腸到達率亦是一項費時且費力的工作，因此，為了更有效地監測盲腸到達率，我們的實驗室成員先前已提出一套盲腸影像辨識的演算法，此演算法在辨識影像上有良好的表現，但在以病例為單位，計算盲腸到達率的表現尚有改善空間。本論文於固有的盲腸影像辨識模型，提出前處理與後處理的方法，其中前處理包含影像裁切以及利用色彩空間分析過濾清腸不佳之大腸鏡影像，後處理則是根據臨床實際情況調整預測結果，進一步改善預測準確率，自動化判斷每一筆病例是否成功到達盲腸，以改善大腸鏡手術的品質。我們提出的方法與之前的結果相比，在內部測試集上，準確率從87.27%提升到91.73%、靈敏度從89.43%提升到91.89%、特異度從85.14%提升到91.57%；在外部測試集上，準確率從75.58%提升到87.59%、靈敏度從89.70%提升到91.60%、特異度從62.46%提升到83.88%，此實驗結果也證明了分類器過適(overfitting)的問題得到改善。

關鍵字

盲腸 ； 盲腸到達率 ； 大腸鏡影像 ； 影像辨識 ； 分類器 ； 影像前處理 ； 色彩空間 ； 後處理 ； 基於知識過濾法

並列摘要

Colonoscopy is widely used for the diagnosis of colorectal cancer (CRC). Lower quality colonoscopy would not only result in missed detection of precancerous lesions, but also easily leads to discomfort and complications. Therefore, the quality of colonoscopy is important. The indicators for evaluating colonoscopy include cecal intubation rate (CIR), adenoma detection rate (ADR), bowel preparation (BP) and withdrawal time (WT). In view of the current situation in Taiwan, whether the colonoscopy has reached the cecum is still based on the endoscopist’s declaration. In addition to the lack of an objective evaluation method, calculating CIR is also a time-consuming and laborious task. Therefore, in order to monitor cecal intubation rate more effectively, our laboratory members have previously proposed a cecum image recognition system. There is a good performance with image-based in the system, but there is room for improvement in the performance of calculating cecal intubation rate with case-based. This thesis proposes pre-processing and post-processing methods based on the inherent cecum image recognition system. Compared with previous work, our proposed method improves the accuracy from 87.27% to 91.73%, sensitivity from 89.43% to 91.89%, and specificity from 85.14% to 91.57% on the internal test set. Moreover, on the external test set, the accuracy increased from 75.58% to 87.59%, sensitivity from 89.70%. To 91.60%, and specificity from 62.46% to 83.88%, which means that the overfitting problem has been improved.

並列關鍵字

Cecum ； Cecal Incubation Rate ； colonoscopy ； image recognition ； classifier ； image preprocessing ； color space ； postprocessing ； knowledge filtering

參考文獻

[1] R. L. Siegel, K. D. Miller, and A. Jemal, “Cancer statistics, 2019,” CA Cancer J. Clin., vol. 69, no. 1, pp. 7-30, 2019.

[2] J. S Lin, M.A. Piper, L.A. Perdue et al., “Screening for colorectal cancer: Updated evidence report and systematic review for the US Preventive Services Task Force,” JAMA, vol.315, no.23, pp. 2576–2594, 2016.

[3] U. Ladabaum, J. A. Dominitz, C. Kahi et al., “Strategies for Colorectal Cancer Screening,” Gastroenterology, vol. 157, no. 5, pp. 1691-1692 2020.

[4] M. Araghi, I. Soerjomataram, M. Jenkins et al., “Global trends in colorectal cancer mortality: projections to the year 2035,” in International Journal of Cancer, vol. 144, no. 12, pp. 2992-3000, 2018.

[5] S. Ameling, S. Wirth, D. Paulus et al, “Texture-based polyp detection in colonoscopy,” in Bildverarbeitung für die Medizin.Berlin, Germany: Springer, pp. 346-350, 2009.