Classification of wireless capsule endoscopy images for bleeding using deep features fusion

Wireless Capsule Endoscopy (WCE) is used to detect various anomalies in the gastrointestinal (GI) tract like ulcers, bleeding, tumor and polyps in a discrete way. Since, it produces a large number of images, whose examination is done by the doctors, which makes this process tedious, time-consuming and further increases the chances of human errors. Hence, this paper presents a novel automated approach for detection of bleeding in WCE images by ensembling three pre-trained CNNs viz. InceptionV3, ResNet152V2 and InceptionResNetV2 to decrease the time needed for examining the large image dataset and also to enhance the accuracy of the analysis. First, the features are extracted by using transfer learning on these pre-trained CNNs. Then Principal Component Analysis (PCA) is used to select the set of optimal features from the set of extracted features. Finally, the selected features are fused and passed to the Support Vector Machine (SVM) to classify WCE images into two categories: bleeding and normal. The proposed approach is applied to 912 WCE images collected from the Department of Gastroenterology, in the University Hospital of Coimbra, Portugal. We have given comprehensive experimental results and a comparative analysis of the proposed approach with other state-of-the-art methods.