Knee Osteoarthritis Detection Using Bone Distances

Knee osteoarthritis (OA) is one of the most prevalent chronic diseases in the United States. The cartilage thickness between the femur and tibia bones is a critical biomarker to diagnose knee OA from screening medical images of the knee. This work aims to study an alternative feature, i.e., bone distance between the femur and tibia in knee MRI, as an alternative to cartilage thickness for knee OA detection. To accurately measure the distance, we first apply a U-Net model to segment the bone regions from the 3D MRI images. Different distance measurement algorithms are proposed to extract features from the segmented images. The dataset contains 3D MRIs of 195 knees from the public OAI database. Each knee is labeled as positive or negative, where positive means OA and negative means healthy knee. Multiple machine learning models (Support Vector Machine, Multilayer Perceptron’s, Random Forest, LightGBM, and Residual Neural Network with Attention Block (RNNAB) were trained to perform the classification task using the distance features. The best performance was achieved using the average distance features in three equally divided regions plus the RNNAB, with an F1 score of 76% and an AUC of 0.76. All the results are averaged over 10-fold cross-validation on the whole dataset. Experiment results show that distance-based features from knee MRI have good potential to diagnose knee OA when coupled with a Residue Neural Network with Attention Block.