Bridging the Resolution Gap between Endoscope Types for a Colonic Polyp Classification

In this work we investigate whether cross endoscopic modality classification of colonic polyps is feasible, i.e. images of a high-magnification endoscope are used as training set to classify images of a high-definition endoscope. In order to compensate the scale differences between the images acquired with the different imaging modalities we apply different super-resolution methods to endoscopic high-definition sequences. We then use a set of feature extraction methods for the classification of the super-resolution reconstruction results. To be able to assess whether super-resolution algorithms are helpful in this scenario, we also compare the results obtained from these experiments against the classification results based on original high-definition frames and against classification rates based on upscaled versions of high-definition frames. We show that classifying images acquired with a high-definition endoscope and training the underlying classifier with images acquired with a high-magnification endoscope is feasible, but the improvements by using super-resolution algorithms are highly feature-dependent.