Performance evaluation of some textural features for muscle tissue classification

Abstract

Textural features are compared for the classification of MR muscle images. The objective is to determine which features optimize classification rate using small ROIs. Four classes of textural features are considered: the authors have studied fractal, cooccurrence, higher order statistics and mathematical morphology. The quantitative evaluation of the discrimination power of the features is based on the performance of the classification error rate with a K-nearest neighbor classifier. The results shows that the mathematical morphology features provide the best classification rate on the authors' clinical MR images of healthy and sick muscles.<>