Weakly-Supervised Lesion Detection in Video Capsule Endoscopy Based on a Bag-of-Colour Features Model

Robotic video capsule endoscopy (VCE) is a rapidly evolving medical imaging technology enabling more thorough examination and treatment of the gastrointestinal tract than conventional endoscopy technologies. Despite of the technological advances in this field, the reviewing of the large VCE image sequences remains manual and challenges experts' diagnostic capabilities. Video reviewing systems for automated lesion detection are still under investigation. Most of these systems are based on supervised machine learning algorithms, which require a training set of images, manually annotated by the experts to indicate which pixels correspond to lesions. In this paper, we investigate a weakly-supervised approach for lesion detection, which requires image-level instead of pixel-level annotations for training. Such an approach offers a considerable advantage with respect to the efficiency of the annotation process. It is based on state-of-the-art colour features, which, in this study, are extended according to the bag-of-visual-words model. The area under receiver operating characteristic achieved, reaches 81%.