Computer-aided Detection Of Hookworm And Most Commonly Seen Contents In Capsule Endoscopy Images

Gastrointestinal(GI)tract disease is a common type of disease in humans.Nowadays,capsule endoscopy(CE)has become the main means of digestive tract inspection as a new technology with the characteristics of safety,convenience,and needless of sedation.After swallowed by a patient,the capsule endoscopy will move through and record about 50 thousand images.It is a tedious and time-consuming task for a clinician to examine massive images.Therefore,a computer-aided diagnosis system is desired and urgent to conduct this task.In this dissertation,we study the detection of hookworms,organs,most commonly seen contents,as well as roundworms in digestive tract based on capsule endoscopic images.This dissertation include four aspects as follows:Firstly,as one of the human helminths,the hookworm bring about serious threatens to the health of Chinese.The computer-aided detection of hookworm through capsule endoscopy images is first proposed in this dissertation by hybrid color gradient map and Contourlet transform.Then a group of color channels from different color mode is defined as a hybrid color mode.In order to capture the edge and texture information of hookworm,the oriented energy filters are applied to each channel of hybrid color mode and then a group of gradient map is produced in which the mucosa texture are suppressed.Following,the hybrid color gradient maps are applied with Contourlet transform,and the edge and texture feature of hookworm are decomposed into coefficiens.Finally,the coefficients are compressed by a group of moment values,and hookworm images are detected by a support vector machine.The comparison experiment shows that the proposed method improves the hookworm detection performance.Secondly,the detection of hookworm based on locating and segmentation is proposed in this dissertation.By applying multi-scale dual matched filter and piecewise parallel region detection,the location of the hookworm-like tubular is found and the corresponding region is segmented accurately.In order to distinguish bubble and fold of intestine,the irregular regions are transformed into regular uncurled tubular regions,and then the histogram of average intensity is proposed to represent their properties.Following,an ensemble learning method named Rusboost is deployed to classify the unbalanced example features of hookworm and non-hookworm.Experiments on a diverse and large-scale dataset with 440 thousand CE images demonstrate that the proposed approach achieves a high accuracy,sensitivity,and specificity and outperforms the state-of-the-art methods.In addition,the proposed method shows promising future in the clinical application of the low missed diagnosis rate.What’s more,as detection of organs and commonly seen noisy contents on digestive tract is a fundamental problem of CE image examination,a general cascaded spatial-temporal deep framework is proposed in this dissertation to detect these contents.The noisy contents such as feces,bile,bubble,and low power images are removed by a Convolutional Neural Network(CNN)model.The rest clear images are then divided into different organs such as the entrance,stomach,small intestine,and colon by a second CNN.In the end,organs are detected accurately by a global temporal integration strategy by Hidden Markov Model.Compared to the existing methods,the proposed framework can detect noisy contents and organs simultaneously.Experiments on a data set with 630 thousand images demonstrate the proposed approach achieves a promising performance in terms of effectiveness and efficiency.The proposed method shows promising future in the clinical application.At last,the detection of Ascaris with specular reflection highlight and histogram of oriented gradient from capsule endoscopy image is proposed in this dissertation.Based on properties of reflection and morphology of Ascaris,the reflection areas on an image are detected with the theory of dichromatic reflection model,and the light-reflecting noisy points of mucosa can be removed.Then,the Ascaris-like area can be detected with connected areas by minimum boundary rectangle.In the end,the Ascaris is detected by the classifier and the histogram of oriented gradient.Experiments show that the proposed method achieved good performances on Ascaris detection without missed patient.