Poster abstract details

Purpose: To assess the possible utility of machine learning for classifying subjects with and subjects without osteoarthritis using sodium magnetic resonance imaging data. Theory: support vector machine, k-nearest neighbors, naive Bayes, discriminant analysis, linear regression, logistic regression, neural networks, decision tree, and tree bagging were tested. Methods: Sodium magnetic resonance imaging with and without fluid suppression by inversion recovery was acquired on the knee cartilage of 19 controls and 28 osteoarthritis patients. Sodium concentrations were measured in regions of interests in the knee for both acquisitions. Mean (MEAN) and standard deviation (STD) of these concentrations were measured in each regions of interest, and the minimum, maximum, and mean of these two measurements were calculated over all regions of interests for each subject. The resulting 12 variables per subject were used as predictors for classification. Results: Either Min [STD] alone, or in combination with Mean [MEAN] or Min [MEAN], all from fluid suppressed data, were the best predictors with an accuracy $>74\%$, mainly with linear logistic regression and linear support vector machine. Other good classifiers include discriminant analysis, linear regression, and naive Bayes. Conclusion: Machine learning is a promising technique for classifying osteoarthritis patients and controls from sodium magnetic resonance imaging data.