Author: Yael Tudela
Supervisors: Jorge Bernal
Presentation time: 10:30
Virtual Room: 3.3 | Live presentation URL
Colorectal cancer is one of the main causes of deaths worldwide. Early detection and histological class prediction of its precursor lesion, the polyp, is key to reduce its mortality and to improve the efficiency of the procedure. In this work we study how two-stage object detection architectures can be used to provide an accurate polyp detection and classification on video sequences while keeping a trade-off between performance and resource consumption that allows the use of such a system in the exploration room. We use Faster R-CNN and Mask R-CNN as base architectures and we propose several cost efficient improvements aiming to alleviate some of the drawbacks that they present. Our experiments show promising results and indicate a benefit associated with the use of more different polyps (even if we only have a few shots of each of them) rather than including additional video sequences.