Automated Detection of Knee Cartilage Region in X-ray Image

Abstract

The prevalence of a symptomatic knee or osteoarthritis (OA) is approximately 9.6% in men and 18.0% in women over 60 years of age according to the OARSI 2016 report. Using early on-stage clinical qualitative assessments through means of X-ray scans, the cartilage health and degradation of an individual can be monitored through cartilage shape and surface over time. In this paper, we implement the application of transfer learning models such as InceptionV3, Xception and DenseNet201 for feature extraction of a rebalanced 1,000 knee X-ray images taken from Osteoarthritis Initiative (OAI) dataset with 5 classes graded 0–4 according to Kellgren-Lawrence grading split into a 70/15/15 training/validation/testing split. The features extracted are subsequently fed into machine learning classifiers, namely support vector machine (SVM). An average multiclass accuracy of 71.33% was achieved for hyperparameter fine-tuned DenseNet201-SVM model.