Visual Odometer on Videos of Endoscopic Capsules (VOVEC)

• Autor(a) principal: Gil Martins Pinheiro
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/113958
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

Since its introduction in 2001, capsule endoscopy has become the leading screening method for the small bowel - a region not easily accessible with traditional endoscopy techniques - revolutionizing the way diagnostics work in the field of small bowel diseases. These capsules are vitamin-sized and leverage from a small wireless camera to create 8 to 10 hour videos of the patients digestive tract. Due to the long duration of the videos produced, the human-based diagnosis is elongated, tedious and error-prone. Moreover, once a lesion is found, the localization information is scarce and hardware dependent, entailing desirability for a software-only endoscopic capsule localization system with added precision. This work stems from this need and, bearing this in mind, we propose the implementation of two deep-learning based methods to improve upon the limitations of the techniques used so far for the capsule position estimation. To train and test our networks, a dataset of 111 PillCam SB3 and 338 PillCam SB2 videos were used, courtesy of Centro Hospitalar do Porto (CHP). The first method consists in a simple capsule displacement estimation throughout the small bowel utilizing HomographyNet, a deep learning supervised approach that is used for homography computation between images. (DeTone et al. (2016)) Differently, the second proposed method is intended to provide a 3D position along the small intestine, utilizing a deep learning unsupervised approach labeled SfMLearner, which takes advantage of a combination between a DepthNet and a PoseNet to learn depth and ego-motion from video simultaneously. (Zhou et al. (2017))