Quantitative comparison of segmentation methods for in-body images

In this paper, we present a numerical comparison of how well segmentation algorithms approximate the manual segmentation of gastroenterologists for a set of endoscopic images. Different areas in these images demand different levels of analysis by a clinician and some provide critical information about the patient. Our objective is thus to segment endoscopic images so that the results mimic as closely as possible the areas that were considered relevant by doctors. We focus on a detailed quantitative comparison of two popular segmentation algorithms, mean shift and normalized cuts, when applied to in-body images, most specifically for vital-stained magnification endoscopy. Segmentation results are compared with the manual annotations of the same images performed by two specialist clinicians. Results show that if we simply consider the most relevant segmented patch, normalized cuts performs better. However, if we allow the annotated area to be represented by multiple patches, mean shift is clearly a better choice, although automatic ways to determine its kernel's bandwidth are highly desirable.