Learning osteoarthritis imaging biomarkers using Laplacian eigenmap embeddings with data from the OAI

We propose a data-driven approach to learn diagnostic imaging biomarkers of osteoarthritis (OA) using multiple regions of the articular cartilage in knee MRI. We discover novel biomarkers for OA diagnosis by learning Laplacian eigenmap manifold embeddings for different regions of interest (ROIs). All embeddings are learnt using MR images from 1131 subjects from the OAI dataset. We show that combining embeddings learnt from different ROIs has better discriminative performance when compared with an embedding constructed using a single ROI. The learnt manifold co-ordinates can be used as biomarkers. The efficacy of the novel biomarkers is tested using Linear Discriminant Analysis (LDA), which linearly projects the diagnostic biomarkers onto a discriminant hyperplane. The area under the receiver-operator curve (AUC) for the diagnostic biomarker is 0.904 (95% confidence interval 0.887-0.920). The results demonstrate that these techniques improve upon results reported in the literature.