Getting one step closer to fully automatized celiac disease diagnosis

Up to now, for computer aided celiac disease diagnosis reliable subimages showing discriminative features must be manually extracted by the physicians, prior to the automatized classification. This must be done to get idealistic data which is free from image degradations, in order to enable a reliable computer based classification. However, this interactive stage during medical treatment requires significant time and attention of the physical doctor. Furthermore, an inadequate selection (e.g. of an inexperienced doctor) leads to a decreased classification accuracy. In this work, a method is proposed to select reliable subimages from the original endoscopic images by maximizing a quality measure. Therefore, for the specific problem definition, we introduce five measures which are supposed to be appropriate for reflecting the adequateness of a subimage, with respect to a specific degradation type. Moreover, as none of the single metrics is able to reflect all prevalent degradations, we propose a weighted combination of these metrics. Extensive experiments have been done with five feature extraction techniques, that turned out to be appropriate for celiac disease diagnosis. Finally the best accuracies are achieved by the metric based on the weighted combination.