Statistical analysis of the impact of distortion (correction) on an automated classification of celiac disease

In this work we investigate the impact of barrel-type distortions and distortion correction on an automated classification of endoscopic imagery. For this purpose we use a set of methods from earlier work along with the nearest neighbor classifier on endoscopic images which are distorted in a barrel-type fashion. In addition we classify images after applying distortion correction to them. Depending on the distance of a patch to the center of distortion a patch is more or less distorted. Hence, we analyze the impact of this distance on the classification accuracy and show that it has an influence on the classification accuracy. We further investigate whether there is a relationship between the levels of distortion of a patch and the respective nearest neighbor patch. We show that a patch and the respective nearest neighbor patch rather often exhibit the same amount of distortion. We further show that higher distortion differences also have an impact on the probability of misclassification for a patch. The results indicate that for higher distortion differences a patch is more likely to be misclassified.