Knee Osteoarthritis Detection and Classification Using X-Rays

Abstract

Knee osteoarthritis is a common form of arthritis, a chronic and progressive disease recognized by joint space narrowing, osteophyte formation, sclerosis, and bone deformity that can be observed using radiographs. Radiography is regarded as the gold standard and is the cheapest and most readily available modality. X-ray images are graded using Kellgren and Lawrence's (KL) grading scheme according to the order of severity of osteoarthritis from normal to severe. Early detection can help early treatment and hence slows down knee osteoarthritis degeneration. Unfortunately, most of the existing approaches either merge or exclude perplexing grades to improve the performance of their models. This study aims to automatically detect and classify knee osteoarthritis according to the KL grading system for radiographs. We have proposed an automated deep learning-based ordinal classification approach for early diagnosis and grading knee osteoarthritis using a single posteroanterior standing knee x-ray image. An Osteoarthritis Initiative(OAI) based dataset of knee joint X-ray images is chosen for this study. The dataset was split into the training, testing, and validation set with a 7: 2: 1 ratio. We took advantage of transfer learning and fine-tuned ResNet-34, VGG-19, DenseNet 121, and DenseNet 161 and joined them in an ensemble to improve the model's overall performance. Our method has shown promising results by obtaining 98% overall accuracy and 0.99 Quadratic Weighted Kappa with a 95% confidence interval. Also, accuracy per KL grade is significantly improved. Furthermore, our methods outperform state-of-the-art automated methods.