Detection of Intestinal Bleeding in Wireless Capsule Endoscopy using Machine Learning Techniques
Date
2020-05-21
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
0000-0002-0609-7670
Type
Thesis
Degree Level
Masters
Abstract
Gastrointestinal (GI) bleeding is very common in humans, which may lead to fatal consequences. GI bleeding can usually be identified using a flexible wired endoscope. In 2001, a newer diagnostic tool, wireless capsule endoscopy (WCE) was introduced. It is a swallow-able capsule-shaped device with a camera that captures thousands of color images and wirelessly sends those back to a data recorder. After that, the physicians analyze those images in order to identify any GI abnormalities. But it takes a longer screening time which may increase the danger of the patients in emergency cases. It is therefore necessary to use a real-time detection tool to identify bleeding in the GI tract.
Each material has its own spectral ‘signature’ which shows distinct characteristics in specific wavelength of light [33]. Therefore, by evaluating the optical characteristics, the presence of blood can be detected. In the study, three main hardware designs were presented: one using a two-wavelength based optical sensor and others using two six-wavelength based spectral sensors with AS7262 and AS7263 chips respectively to determine the optical characteristics of the blood and non-blood samples.
The goal of the research is to develop a machine learning model to differentiate blood samples (BS) and non-blood samples (NBS) by exploring their optical properties. In this experiment, 10 levels of crystallized bovine hemoglobin solutions were used as BS and 5 food colors (red, yellow, orange, tan and pink) with different concentrations totaling 25 non-blood samples were used as NBS. These blood and non-blood samples were also combined with pig’s intestine to mimic in-vivo experimental environment. The collected samples were completely separated into training and testing data.
Different spectral features are analyzed to obtain the optical information about the samples. Based on the performance on the selected most significant features of the spectral wavelengths, k-nearest neighbors algorithm (k-NN) is finally chosen for the automated bleeding detection. The proposed k-NN classifier model has been able to distinguish the BS and NBS with an accuracy of 91.54% using two wavelengths features and around 89% using three combined wavelengths features in the visible and near-infrared spectral regions. The research also indicates that it is possible to deploy tiny optical detectors to detect GI bleeding in a WCE system which could eliminate the need of time-consuming image post-processing steps.
Description
Keywords
Wireless capsule endoscopy, gastrointestinal bleeding, optical sensor, feature selection, machine learning
Citation
Degree
Master of Science (M.Sc.)
Department
Electrical and Computer Engineering
Program
Electrical Engineering