基於深度學習之盲腸影像辨識技術應用於醫療資訊系統

Abstract

大腸癌是近年來世界上一個重要的問題。受苦的人數和醫療費用的成本仍在增加。為了預防大腸癌，定期檢查和早期治療是最有效的方法。因此，標準和高品質的大腸鏡檢查是非常重要的。因此在本論文中，我們專注於大腸鏡檢查的其中一項重要指標：盲腸到達率。 鑑於台灣目前的情況，大腸鏡檢查是否到達盲腸仍然是基於醫生的聲明，沒有客觀的評價方法。目前的內視鏡醫師難以有額外的人力來檢查每個大腸鏡檢查影像是否到達盲腸且大腸鏡檢查影像的品質是否符合標準。為了更有效地監測盲腸到達率和照片品質，因此提出了一種自動識別盲腸的系統。醫生將大腸鏡檢查圖像上傳到該系統後，系統可以將圖像自動區分為盲腸或非盲腸，以實現盲腸到達率的自動計算。希望這種自動化模式可以節省人力讓醫務人員可以花更多時間 在病人身上。 該系統基於我們實驗室先前研究的圖像分析方法。它首先評估大腸鏡檢查的腸道準備並區分腸道是否乾淨與否。乾淨的影像才會進行預處理和資料擴增。接著使用卷積神經網絡演算法，如GoogLenet ，VGGNet網絡和其他著名的神經網絡架構最終選擇了最好的網絡作為我們的模型。除了來自台大醫院大量的大腸鏡檢查圖像。我們另外使用GPU來加速訓練，然後讓計算機自動學習盲腸與非盲腸圖像的差異。另外為了知道機器學習到甚麼，因此讓機器呈現以視覺上可理解的方式識別特徵，然後識別影像是否已到達盲腸，最後統計各項數據後將病人相關資料以及辨識結果寫入MSSQL資料庫，使用網頁平台方式來呈現相關統計數據。 此外，為了促進這種管理系統遍布全國各大醫院，我們還與醫生合作，收集了NTU醫院以外的其他10家醫院的大腸鏡檢查照片。它可以增加照片的多樣性，增強我們系統的穩健性。因為可以應用大量的照片作遷移式學習來突破準確性的瓶頸並最終通過接收器操作特性曲線，選擇最佳閾值來獲得優化模型。在未來，該系統有望幫助醫生確定他們是否真的在大腸鏡檢查期間進入盲腸， 作為第三方公正評大腸鏡檢查的品質，同時減輕醫生肉眼觀看照片的負擔。

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)。 In view of the current situation in Taiwan, whether the colonoscopy reaches the cecum is still based on the physician's declaration and there is no objective evaluation method. It is difficult for the current endoscopic physician to have the extra manpower to examine whether each colonoscopy image reaches the cecum and the quality of the colonoscopy image meets the standard or not. In order to monitor the cecal intubation rate and photo quality more efficiently, a system for automatically identifying the cecum was proposed. After the doctor uploads the colonoscopy image to the system, the system can distinguish the image as a cecum or a non-cecum to achieve automatic calculation of the cecal intubation rate. It is hoped that this automated mode can save manpower and allow the medical staff to spend more time on the patient. This system is based on the image analysis method previously study of our lab. It first evaluates the bowel preparation of the colonoscopy and distinguishes whether the bowel is clean or not. Image pre-processing and data augmentation are implemented on clean images. Next use the convolutional neural network algorithm such as GoogLenet VGG, Residual Network and other well-known neural network architectures finally select the best network architecture as our model. In addition to this the large amount of colonoscopy images from the NTU hospital are also needed .We use the GPU to accelerate the computer train the model then let the computer automatically learn the features differences of the cecum and non-cecum images. We also want to know what machine learns so we want to present the features in a visually understandable way then identify whether the photo has been photographed into the cecum or not. Finally, the patient-related data and the photo classification result are written into the MSSQL database, and the web platform is used to present relevant statistics. In addition, in order to promote this system spread across major hospitals across the country, we also collaborated with doctors to collect colonoscopy photos from 10 other hospitals other than the NTU Hospital. It can increase the diversity of photos and enhancing the robustness of our system. Because large amount of photos transfer learning can be applied to break through the bottleneck of the accuracy and finally through the receiver operating characteristic curve, select the optimal threshold to get the optimization model. In the future, this system is expected to assist doctors in determining whether they have actually entered the cecum during colonoscopy as a third party to fairly assess the quality of colonoscopy, while reducing the burden of manual viewing of photos.