Machine learning for automatic grading of knee osteoarthritis from X-ray radiographs

Knee osteoarthritis is a growing problem due to increasing risk factors such as age and obesity. It is a common task for a radiologist to grade osteoarthritis in three compartments (medial tibiofemoral (MTF), lateral tibiofemoral (LTF) and patellofemoral (PF)) in a knee from different image views of X-ray images, to decide if osteoarthritis is the cause of pain for the patient. Reasons for automating this process are to decrease subjectivity, time for diagnosis and reduce workload for radiologists. The aim with this project was to grade osteoarthritis using machine learning by training convolutional neural networks on around 5000 double annotated examinations by radiologists and one orthopaedic surgeon at Nyköping Hospital. Different methods were evaluated and the models were then optimised with hyperparameter tuning. The aim with the project is to contribute to a future software that could be tested at Nyköping Hospital. The project found that using transfer learning with DenseNet for MTF and PF, and using a MTF model as transfer learning model for the LTF model was the best performing transfer learning networks to use. Also, cropping the images around the region of interest for MTF and LTF improved the models. The best method to make predictions from the model outputs appeared to be to train a model on a merged set of training- and validation data for making predictions. Comparisons of final models with the radiologist initial annotations showed that the MTF and LTF models give fewer misclassifications of more than one grade, if compared to the disagreements of more than one grade by the two radiologists. While for the PF model the radiologists still have an advantage and more data is probably needed for both the PF model and the LTF model since grade 0 is very overrepresented for those grades