An expert system for lesion detection in wireless capsule endoscopy using transfer learning

Wireless Capsule Endoscopic (WCE) is a powerful diagnostic tool that has proven especially useful in imaging the small intestine. Currently its use is limited due to the production of a considerable number of images whose analysis is an extremely time-consuming process. The present work presents the design and implementation of a novel computer-aided diagnosis system for automatic classification of images acquired by the capsule into image with and without lesion. Our expert system is based on the concept of transfer learning which reuses features of pre-trained neural networks needing little data. Two pre-trained deep convolutional neural (ResNet50 and Inception V4) were used for feature extraction. Then, the selected features are combined using the minimum Redundancy Maximum Relevance technique. For the final classifier layer, specific machine learning classifiers that have shown promising results in previous medical images studies were compared. The experiments performed on two standard benchmark datasets demonstrated that our expert system outclass the single deep learning architectures, with an average accuracy in detection lesions of 98.09 % on KID Dataset 2 and 94.48 % on MICCAI 2017. The best results in terms of accuracy and training time were obtained using Support Vector Machine as classifier.