Region of Interest Segmentation in Histopathological Images of Colorectal Polyps

Histopathological images often contain a lot of diagnostically irrelevant, distracting information. The pathologist needs to focus on specific regions where he can observe details as well as the shape and number of larger cellular structures. In this paper, we present two approaches to labelling regions of interest and learning segmentation models for automatic detection of these regions. The first approach was so-called coarse labelling, which is less laborious and more time-efficient for the labeller. In this experiment, 123 images were labelled. It turned out that the segmentation model trained on this data was more accurate than the labels themselves. The second approach was the so-called fine labelling, which is much more time-consuming for the labeller. Only 10 images were labelled using this method. Despite the extremely small training data set, the model trained with this data segmented the regions of interest better than the model trained with coarse labels.