Automatic Multi-Feature Extraction for Clinical Colonoscopy Image Analysis and Recognition

We proposed an analysis system to extract the features of cecum and non-cecum structures of clinical colonoscopy images. We found out the different expression of these two cases and obtained multiple features after extraction as the training data for our classifier of system. The proposed algorithm helped improve the accuracy and robustness compared to the previous work. Besides, we also provided assistance to doctors for their judgments with the result of image processing on colonoscopy images. In recent years, the dietary habit and life of people had become very different nowadays, and the concerns of the digestive system related diseases had become more and more serious issue to people. Among these, colorectal cancer has the high incident rate and high cost within the top ranks in our country. To prevent colorectal cancer and maintain good digestive environment, the demand of regular colonoscopy is increasing around the world. However, to achieve the purpose of the examination effectively, complete procedure is essential. The cecal intubation rate (CIR) is the critical quality indicator of colonoscopy completion, which is our system mainly focus on. We proposed an image processing based system, which could automatically recognize the landmarks of cecum images such as Ileocecal valve (ICV), Tri-radiate Fold and Appendiceal Orifice. In addition, we also found the structures of non-cecum images. And we determined the features to demonstrate the difference between these two cases performance for our classification accordance. We developed an approach to analyze colonoscopy images, including reducing the noises such as reflection, veins and dirt of excrements. Further, we extracted vector of multiple features to describe cecum and non-cecum structures’ performance properly for the classification and recognition in our system. In this thesis, we applied Particle Filter algorithm to have better result compared to previous work, which used merely single lightness threshold as the classification feature. Moreover, Edge-Based, Histogram-Based and Texture-Based are the additional aspects we utilized to extract our features for colonoscopy images analysis, including classification and segmentation. Finally, we pass our feature vector as training data to another partner in our project, who is focus on developing Machine Learning using Adaboost algorithm for the classification dealing with the photos uploaded by doctors. The material of our experiment was provided by NTU hospital, which were totally 664 patients’ folder of clinical colonoscopy images. We randomly selected from it and got 1009 cecum images and 997 non-cecum images for our training and testing purpose. At last, we not only got average 94.0% and best 96.9% accuracy of classification performance, but also marked the vital landmark of cecum and non-cecum structures for doctors’ analysis.