Automatic Bleeding Detection in Wireless Capsule Endoscopy Images

Abstract

Image segmentation in medical images is performed to extract valuable information from the images by concentrating on the region of interest. Mostly, the number of medical images generated from a diagnosis is large and not ideal to treat with traditional ways of segmentation using machine learning models due to their numerous and complex features. To obtain crucial features from this large set of images, deep learning is a good choice over traditional machine learning algorithms. Wireless capsule endoscopy images comprise normal and sick frames and often suffers with a big data imbalance ratio which is sometimes 1000:1 for normal and sick classes. They are also special type of confounding images due to movement of the (capsule) camera, organs and variations in luminance to capture the site texture inside the body. So, we have proposed an automatic deep learning model based to detect bleeding frames out of the WCE images. The proposed model is based on Convolutional Neural Network (CNN) and its performance is compared with state-of-the-art methods including Logistic Regression, Support Vector Machine, Artificial Neural Network and Random Forest. The proposed model reduces the computational burden by offering the automatic feature extraction. It has promising accuracy with an F1 score of 0.76.