Detection and Classification of Human Colorectal Polyps

In this thesis we address the problem of segmentation and detection of colorectal polyps in endoscopic images obtained by a wireless capsule endoscopic device. As colorectal polyps are precursors to colonic cancer, and an early detection and removal of colorectal polyps can positively affect the prognosis of colonic cancer (one of the leading causes of death by cancer), it is our goal to develop an computer aided diagnosis method to detect colorectal polyps. With this in mind, a segmentation and detection method for colorectal polyps images acquired by wireless capsule endoscopy was developed. The cornerstone of the approach of segmentation and detection is the fact that, as polyps are protrusions emerging from colonic walls, the polyps can be segmented by curvature descriptors. The curvature descriptors require denoised images and their continuous representation. Because of this, we use local estimations of the images (Local Polynomial Approximation) which provide both denoising and a continuous representation of the images. A wrapper feature selection algorithm (greedy forward feature selection algorithm) was implemented in order to determine a set of features for the classification. The classification is performed using Support Vector Machines, being obtained several sets of features resulting of the progression metrics used, balancing of the training set, and type of kernel for the SVM. The segmentation and detection method using the classifier was tested in several scenarios presenting good results both using the same video sequences as training data and testing data (cross-feature validation) and different video sequences as training and testing data.