Automatic knee cartilage segmentation using fully volumetric convolutional neural networks for evaluation of osteoarthritis

Automated Cartilage segmentation is essential for improving the performance of advanced Knee Osteoarthritis (OA) assessment due to its convoluted 3D structure. In this paper, we have developed a knee cartilage segmentation algorithm from a high resolution MR volume using a novel 3D-fully Convolutional Neural Network (CNN), called `μ-Net' coupled with a multi-class loss function. This is, to our knowledge, the first automatic cartilage segmentation method using 3D CNNs. The proposed algorithm performed better than the state-of-the-art algorithm in the MICCAI SKI 10 public challenge. We have further applied our proposed algorithm on another similar MR contrast (DESS) provided by Osteoarthritis Initiative (OAI) for OA assessment and have presented improved segmentation accuracies. Initial qualitative assessment of segmentation results visually depicts cartilage loss in longitudinal knee MR data.