Modality Classification Using Texture Features

Abstract

Medical image classification based on the image modality is one of the most important and crucial tasks in the medical image analysis. Due to its importance, the aim of the paper is to investigate modality medical image classification problem by using a combination of several classification techniques and feature extraction algorithms over a set of medical images. Four feature extraction methods were used in this paper: LBP, GLDM, GLRLM, Haralick texture features. Additionally we concatenated all four features in one single feature to assess their joint performance. The feature extraction algorithms are tested over three classifiers: SVM extended for multiclass classification based on one against all strategy, k-nearest neighbor and C4.5 algorithm. This examination was conducted over a set of medical images provided by ImageCLEF. The dataset contains 18 classes of images, with a total of 988 images. The distribution of number of images per class is not uniform, so this additionally burdens the task. The best results were provided when the images were described with concatenated descriptors and classified with SVM classifier, with a classification accuracy of 73%.