A Posteriori Deep Learning Segmentation Quality Estimation Based on Prediction Entropy

Abstract

Image segmentation is a common intermediate operation in many image processing applications. On automated systems it is important to evaluate how well it, or its subsystems are performing without access to the Ground Truth. In Deep Learning based image segmentation there are very few methods to evaluate the output quality without using a ground truth. Most of them are based on the uncertainty (variance or standard deviation) of the prediction and can be applied to Bayesian Neural Networks, but not to Convolutional Neural Networks. In this research we propose to use the Entropy as a measure of uncertainty applied to the segmented image predicted by the Neural Network and some indicators based on it. The method is tested in a segmentation task of labeled skin images. The entropy based indicators are evaluated without knowing the ground truth and compared with indicators based on the real labels (Jaccard, Dice and Average Symmetrical Surface Distance). This experimentation showed that they are correlated and some Entropy based indicators can predict quite well the ground truth based indicators.