Wireless Capsule Endoscopy Bleeding Images Classification Using CNN Based Model

Abstract

Wireless capsule endoscopy (WCE) is an efficient tool to investigate gastrointestinal tract disorders and perform painless imaging of the intestine. Despite that, several concerns make its wide applicability and adaptation challenging like efficacy, tolerance, safety, and performance. Besides, automatic analysis of the WCE provided dataset is of great importance for detecting abnormalities. Imaging of the patient's digestive tract through WCE produces a large dataset that requires a substantial amount of time and a special skill set from a medical practitioner for analysis. Several computer-aided and vision-based solutions have been proposed to resolve these issues, yet, they do not provide the desired level of accuracy and further improvements are still needed. The current study aims to devise a system that can perform the task of automatic analysis of WCE images to identify abnormalities and assist practitioners for robust diagnosis. This study adopts a deep neural network approach and proposes a model name BIR (bleedy image recognizer) that combines the MobileNet with a custom-built convolutional neural network (CNN) model to classify WCE bleedy images. BIR uses the MobileNet model for initial-level computation for its lower computation power requirement and subsequently the output is fed to the CNN for further processing. A dataset of 1650 WCE images is used to train and test the model using the measures of accuracy, precision, recall, F1 score, and Cohen's kappa to evaluate the performance of the BIR. Results indicate the promising outcomes with achieved accuracy, precision, recall, F1 score, and Cohen's kappa of 0.993, 1.000, 0.994, 0.997, and 0.995 respectively. The accuracy of the BIR model is 0.978 with the Google collected WCE image dataset which is better than the state-of-art approaches.