gms | German Medical Science

Purpose: One of the most common and in the frequency increasing diseases in the western world is the gastroesophageal reflux disease (GERD). Intestinal metaplasia, also known as “Barrett’s Esophagus” is a precancerous condition and complication of GERD. Unfortunately, for the histopathology of Barrett’s Esophagus, a large inter-observer variation can be observed. Hence, a quantitative image-driven approach of computer assisted diagnosis (CAD) to support pathologists with an automated pre-analysis of the images is considered. Goal of this contribution is the evaluation of possibilities to differentiate histopathology automatically three types of tissue, namely normal squamous epithelium of the esophagus (EP), normal cardia mucosa (CA) and Barrett’s Esophagus (BE).

Material and Methods: From 86 randomized selected patients with Barrett’s Esophagus histological slides have been digitized with a high-resolution whole-slide scanner (3DHistech) and the images were anonymized. 26 data sets were selected in which all three tissue classes (BE, CA, EP) are depicted equally. In these data sets, for each class 50 rectangular regions were labeled manually, and then automatically split in square tiles in various scales (1024x1024, 512x512, 256x256). Only tiles depicting at least 60% of the related tissue were considered. Depending on the scale, different amounts of image tiles were available, namely 143 tiles in the 1st (1024x1024) scale (BE 50, CA 40, EP 53), 468 tiles in the 2nd (256x256) scale (BE 137, CA 124, EP 207), and 1914 tiles in the 3rd (128x128) scale (BE 526, CA 548, EP 840). It was evaluated, which image type of image-based color, structure and texture features can be used to the best differentiation of the tissue classes, namely color enhanced 2nd order texture statistics and statistical-geometrical features. Furthermore various parameters and combinations for the texture features were evaluated. For the classification step a nearest-neighbor-classifier with various parameter settings and n-fold cross-validation was applied. For each experiment all classification rates as well as the confusion tables were computed.

Results: Based on a combination of 2nd order statistical texture features, a maximum diagnostic classification rate of 94% (BE 96%, CA 90%, EP 100%) could be achieved on an image data set of raw-data 1024x1024 tiles, which were furthermore reduced to a quarter-size using a Gaussian pyramid. These results denote the possibility of a correct differentiation of the diagnostic classes with respect to the annotated ground truth. The highest possible classification rate for the discrimination of Barrett’s Esophagus was achieved with 96% on a basis of color-extended sum- and difference histograms and correlates to a total classification rate over all three classes of 94% (CA 85%, EP 96%).

Discussion: These results of show that the classes evaluated (BE, CA, EP) can be differentiated on this image data base by applying color-extended texture features, whereas the detection and elimination of EP tissue is possible with a high sensitivity. Hence, the basic criteria have been defined on which experiments can be made in order to differentiate and pre-sort various types of tissue of the esophagus in histological recordings using image analysis methods and possibly detecting conspicuous tissue automatically.