Machine Learning Approaches for the Classification of Knee Osteoarthritis

Abstract

Machine learning systems have gained tremendous attention in solving complex tasks. For instance, image classification using machine learning has multiple applications in the medical field, where different imaging modalities are used for diagnosing, detecting, and classifying diseases. This study aims to analyze different machine-learning approaches for detecting knee osteoarthritis using X-rays. We used different machine learning algorithms to classify X-ray images of the knee to various levels of osteoarthritis severity. We have solved the problem as a binary classification task, multi-class classification, and ordinal regression task. The Histogram of Oriented Gradients and Haralick Features are used for feature extraction in the binary classification task. For Binary Classification, six experiments are performed using different machine learning classifiers. Logistic regression has a maximum accuracy of 84.50%, F measure of 81%, AUC of 0.84, Mathew’s Correlation Coefficient of 0.69, and Cohen’s Kappa 0.66. 10-Fold cross-validation also reveals that the models’ performance is consistent across all folds. Furthermore, texture and shape features are extracted for ordinal classification using the histogram of oriented gradients and dense DAISY descriptors. Finally, separate experiments are performed, and results for both descriptors are compared. Our implemented approaches have shown state-of-the-art results.