Automatic Grading of Knee Osteoarthritis from Plain Radiographs Using Densely Connected Convolutional Networks

In this paper, we consider densely connected convolutional networks and their applicability to the problem of assessment of knee osteoarthritis (OA) severity in the five-point Kellgren-Lawrence scale. First, we use trained from scratch Single Shot Detector (SSD) to localize knee joint areas in radiographs. Then, we apply DenseNets to quantify OA stages in the images of detected knee joints. We consider networks of different depths, trained both from scratch and pre-trained on the ImageNet dataset and fine-tuned in the images from Osteoarthritis Initiative dataset (OAI). Also, different loss functions are examined to understand which one gives the best training results. In the knee joint localization task, we obtain an accuracy of 94.03% under the Jaccard index threshold of 0.75. Also, our classifier outperforms the current state-of-the-art with accuracy of 71% in the classification task.