Color instance segmentation and classification of cervix images

Abstract

Instance segmentation is the task of assigning a label to each pixel in the image, treating objects of same class separately. On the other hand, classification assigns a label to the whole image. In this paper, we are comparing both methods on a data-set of different cervix types. It was required to detect which cervix type out of 3 categories the image holds. For classification using Convolutional neural network, the region of interest is segmented before starting to train the network. However, in instance segmentation, the input is the full image.Instance segmentation happened to outperform the classification pipeline on this data-set with accuracy of 62% vs 55% for the latter approach.