Dealing with Intra-Class and Intra-Image Variations in Automatic Celiac Disease Diagnosis

Computer aided celiac disease diagnosis is based on endoscopic images showing the villi structure in regions of the small bowel. Especially unavoidably variable illuminations and varying viewing angles of the individual villi are a source for high intra-class as well as intraimage variations in the image domain. We clarify that common texture descriptors are unable to compensate such a high degree of variance, which is supposed to be a crucial problem in computer aided diagnosis. In this work, a straight-forward split and merge approach is presented which facilitates the final classification task by reducing the intra-image variance and simultaneously enlarging the training set. Using different well known feature extraction techniques as well as two classifiers, it can be shown that the overall classification accuracies can be increased consistently. Additionally, the proposed approach is compared to the related but more complex bag-of-visual-words method.