Boosting Small-Data Performance of LBP: A Case Study in Celiac Disease Diagnosis

A major issue in computer aided celiac disease diagnosis is the prevalence of substantial intra-class and even intra-image variations. A method which splits the images into a set of smaller ones and finally applies a decision level fusion turned out to be a powerful technique to address these problems and to boost the classification accuracy. This is especially true if using Local Binary Patterns and derivatives. However, due to the sparsity and roughness of the final feature vectors, these methods are not optimal if being applied to such small images. Therefore, in this work two novel and two methods from literature are investigated to improve the performances of Local Binary Patterns. Experiments show that the overall classification accuracies can be improved, especially by means of a combination of the novel methods. The techniques presented in this work are not restricted to this certain problem definition. They rather can be applied in arbitrary scenarios with small sized image data.