Mining Colorectal Polyp Images for Colon Examination based on Texture Description and Decision Tree Induction

Medical disease examination is often based on images. Mining these images in order to obtain the classification knowledge for automatic image classification is a challenging task. This task belongs to the field of image mining. Image mining is usually not only comprised of mining a table of numbers it has also to do with transforming the image in the right image description. Both, the image description and the classification knowledge, determine the quality of the classifier. Texture is a powerful method to describe the appearance of different biological objects in images. There are different texture descriptors around. Which one is the best for medical images is still an open question. The most used texture descriptor is the well-known Haralick ́s texture descriptor. We propose a texture descriptor based on random sets. This descriptor gives us more freedom in describing different textures. In this paper we develop two classification models based on decision tree induction, one for each of the two texture descriptors. We compare the two texture descriptors based on a medical data set. We review the theory of the two texture descriptors and describe the procedure for the comparison of the two methods. A medical data set is used that is derived from colon examination. Decision tree induction is used to learn a classifier model. Cross-validation is used to calculate the error rate. The comparison of the two texture descriptors is based on the error rate, the properties of the two best classification models, the runtime for the feature calculation, the selected features, and the semantic meaning of the texture descriptors. Mining Colorectal Polyp Images for Colon Examination 97