Detecting Polyps in Endoscopic Images with U-Net Based Architectures - A Preliminary Evaluation

This paper aims to explore some proposed solutions for solving the problem of automatic polyp detection in endoscopic images. These solutions rely on four state of the art Convolutional Neural Networks, designed for the medical image segmentation. The four architectures are the U-Net, U-Net++, ResUNet and ResUNet++, which were tuned by changing the values of different hyperparameters with the aim of achieving the best results. Binary image classification was used to make classifications at pixel level, having two classes: polyp and background. The achieved results were compared using some predefined metrics such as Precision, Recall and mean Intersection over Union. The best results were achieved by the ResUnet architecture, with performances close to those of alternative approaches of the problem.