盲腸影像辨識系統基於深度學習

Cecum Classification of Colonoscopy Images using Deep Learning Algorithm

盧柏諺(Po-Yen Lu)

指導教授 ： 陳中平

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(2017年)

https://doi.org/10.6342/NTU201702857

摘要

在此篇論文中，我們基於深度學習的理論，建立一套自動辨識大腸鏡照片是否具有盲腸的結構，藉此判斷鏡檢醫師是否達成全大腸的掃描，驗證大腸鏡篩檢盲腸到達率能力。
大腸直腸癌不管在台灣、或是全球，都是相當重要的議題。罹患的人數和醫療支出費用更是居高臨下。為了有效降低大腸直腸癌死亡率，定期大腸鏡篩檢與早期治療是最有效的方法。因此，標準且高品質的大腸鏡篩檢是必須的。本論文針對的部分是大腸鏡篩檢品質中的一個指標：盲腸到達率(CIR)。
有鑑於目前在台灣，大腸鏡是否到達盲腸仍以醫師申報為主，缺乏客觀評估方式，以現行內視鏡醫師工作量也難有多餘人力逐一檢視每筆大腸鏡圖片是否到達盲腸。為了更有效率監控盲腸到達率此一品質指標，我們提出一套自動辨識盲腸的系統。醫師將大腸鏡影像上傳此系統後，系統能分辨此影像為盲腸或非盲腸，以達到自動計算盲腸到達率，希望透過此自動化模式能夠不需太多人力就能達成
品質管理的目的。
此系統基於本實驗室先前研究的影像分析方法，先評估大腸鏡照片清腸不潔的程度，分出清腸是否乾淨。再針對清腸較乾淨的照片判斷是否具有盲腸的特徵。我們利用深度學習中的卷積神經網路演算法，利用大量的大腸鏡影像，訓練電腦去學習盲腸與非盲腸影像的特徵與差異，藉此達到辨識照片是否有拍攝到盲腸。最後我們在平均辨識準確率上達到90.14%，以及平均6.77%的無法辨識。未來希望這套系統能協助醫師，判斷是否有確實在大腸鏡檢查中進入盲腸，做為一個公正評估大腸鏡品質的第三方，同時減少人工檢視照片的負擔。

關鍵字

深度學習 ； 卷積神經網路 ； 大腸鏡影像 ； 盲腸辨識 ； 盲腸到達率

並列摘要

In this thesis, based on the theory of deep learning, we establish an automatic cecum recognition system to identify whether a colonoscopy image consists of cecum structure or not. Thereby, this system can check colonoscopist’s real cecal intubation rate to secure good quality of colonoscopy performance on preventing colorectal cancer (CRC).
CRC is an important issue in the world. The number of people suffering and the cost of medical expenses is still increasing. In order to preventing colorectal cancer, regular examination and early treatment is the most effective method. Therefore, standard and high quality colonoscopy screening is necessary. In this thesis, we focus on an indicator of quality colonoscopy: cecal intubation rate (CIR).
Current reporting system regarding to CIR in Taiwan is reported by endoscopist only, no objective external validation system to check the reality about reporting CIR. For inadequate manpower, it is also impossible for endoscopist to check every cecal picture to calculate CIR. For monitoring this quality indicator more efficiently, we propose a recognition system to automatically identify the cecum images. The doctors can upload the images to our platform system, and the system will distinguish between cecum images and non-cecum images to calculate CIR. We hope to achieve good quality control for cecal pictures by this automatically recognition system.
This system is based on the image analysis method of previous study of our lab and deep learning algorithm. Firstly, we evaluate the bowel preparation and pick up the images with clean for recognizing. We use convolutional neuron network in deep learning algorithm and lots of colonoscopy images. We train a model to learn the structure of cecum and non-cecum images and their differences, so we can achieve the goal of cecum identification. Finally, we reach an average accuracy of 90.14%, and an average of 6.77% of the unrecognizable. Also, this system can be a fair assessment of the colonoscopy quality and reducing the burden of manually viewing colonoscopy images.

並列關鍵字

Deep learning ； Convolutional neural networks ； Colonoscopy images ； Cecum recognition ； Cecal intubation rate.

參考文獻

[1] 衛生福利部中央健康保險署,"103年癌症登記報告", http://www.nhi.gov.tw/.

[2] Cunningham D, Atkin W, Lenz HJ et al., "Colorectal cancer," Lancet. 375 (9719): 1030–47. 2010.

[3] Watson AJ, Collins PD., "Colon cancer: a civilization disorder," Digestive Diseases. 29 (2): 222–8. 2011.

[4] 衛生福利部中央健康保險署,"104年各類癌症健保前10大醫療支出統計", http://www.nhi.gov.tw/

[5] Chiu HM, Chen SL, Yen AM et al., "Effectiveness of fecal immunochemical testing in reducing colorectal cancer mortality from the One Million Taiwanese Screening Program," Cancer, 121:3221–9. 2015.