EXTRACTION OF TEXTURE FEATURES FROM KNEE JOINT MRI IMAGES USING MACHINE LEARNING ALGORITHMS FOR CLASSIFICATION OF MENISCAL TEARS

Abstract

This paper presents an approach to automatic classification of knee meniscus condition based on texture analysis of MRI images using machine learning methods. Feature extraction was performed with MaZda 4.6 software, including histogram features, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and wavelet features. The most informative features were selected using RFE, ANOVA, and Fisher methods, reducing the original set from 297 to 16 features. For binary classification of the normal and tear classes, MLP and SVM algorithms were applied, achieving high accuracy (up to 95%) when optimal feature subsets were used. The results confirm the diagnostic significance of texture characteristics and demonstrate their effectiveness for automated detection of meniscal tears, which may be useful for supporting clinical diagnosis and prognostic assessment of knee joint condition.