Improved endoscope distortion correction does not necessarily enhance mucosa-classification based medical decision support systems

Abstract

Distortion correction in two variants is applied to endoscopic duodenal imagery aiming at an improvement of automated classification of celiac disease affected mucosa patches. In a set of heterogeneous feature extraction techniques, only geometry and shape related ones are able to benefit from distortion correction, while for others, even a decrease of classification accuracy is observed. Different types of distortion correction do not lead to significantly different behaviour in the observed application scenario.

Evaluation of different distortion correction methods and interpolation techniques for an automated classification of celiac disease

“…In accordance to the work in [9,11,10] , we consider the four classes Marsh-0 and Marsh-3a to Marsh-3c only throughout this work, since visible changes in the villi structure can be observed only for classes Marsh-3a to Marsh-3c (in case of the classes Marsh-1 the number of intra-epithelial lymphocytes is increased and in case of Marsh-2 the crypts of Lieberkuhn are proliferated). In addition, in this work we focus on the 2-classes case only (i.e.…”

“…The distortion function can then be formulated as with r = || x − x c || 2 and L (·) being a function which uses T ′ to find an appropriate distortion strength for the given r . The implementation used for this method is a custom software, which has already been used in [11,12] .…”

“…To the best of our knowledge up to now there are only a few studies available, which investigate the impact of barrel-type distortions and distortion correction on the accuracy of automated classification systems for endoscopic images [9–12] .…”

“…Gschwandtner et al showed that most feature extraction methods evaluated failed to take advantage of applying distortion correction as a pre-processing step to the endoscopic images, resulting in a decreased classification accuracy. In the follow-up work in [11] the authors evaluated another distortion correction method with a similar outcome. In [12] the authors evaluated the distortion correction methods from [9,11] with additional feature extraction methods.…”

“…In the follow-up work in [11] the authors evaluated another distortion correction method with a similar outcome. In [12] the authors evaluated the distortion correction methods from [9,11] with additional feature extraction methods. In this work the authors showed that only a few feature extraction methods are able to take a slight advantage of distortion correction (with respect to the overall classification rates).…”