Knee Bone Segmentation from MRI Images Using a Deep Learning Model

Abstract

This study proposed an automated segmentation method of knee bone from magnetic resonance imaging (MRI) slides, based on a Deep Learning Convolutional Neural Networks (CNN) architecture. The knee osteoarthritis (OA), oftentimes called “Degenerative Joint Disease” or “Wear and Tear” arthritis, is the most common form of over 100 types of arthritis. The effectiveness of MR images to safely provide a fast 3D visualization of the bones, cartilages, tendons and with good tissue contrast has resulted in its extensive use for the diagnosis and therapy of pathologies of the knee joint. Therefore, the need to accurately segment the knee cartilages and bones from the MR images, which is a challenging process, is paramount to the integrity of the diagnosis and monitoring of the progression of OA pathology.

This study proposed a novel method for the knee bone segmentation using MRI. The proposed method is called Simplified-UNet (S-UNet), which is an improved model based on the original U-Net model. The proposed method achieved Dice coefficient of 95.38%, which is 4.73% better than that of the state-of-the-art U-Net models. The knee MRI images were obtained from the public Osteoarthritis Initiative (OAI) dataset. The ground truth of the bone regions were manually delineated by trained personnel. The dataset contains 88 cases and was divided into training, testing, and validation sets with the ratio of 70%, 15%, and 15% respectively. The improved bone segmentation accuracy on the same dataset shows the superiority of the proposed model over the original U-Net model, indicating the potentials of future application in clinics to assist cartilage segmentation and knee OA diagnosis.