Evaluation of Deep Learning Models in the Classification of Osteoarthritis from Knee X-Ray Image

Abstract

Osteoarthritis (OA) is a common joint disease that is characterized by pain, stiffness, and decreased range of motion over a long time which causes disability and ultimately a poor quality of life. It is important for OA to be diagnosed as early as possible and a good diagnosis can be deemed to have occurred when appropriate therapeutic intervention and management approaches are considered. This work assesses the feasibility of diagnosing OA using knee X-rays with deep learning models including CNN, GoogLeNet, VGG19Net, and AlexNet. Each model was trained on Kaggle’s "Osteoarthritis Prediction" labeled dataset to distinguish between healthy and osteoarthritic joints. Out of the models analyzed, the CNN claimed the highest net classification accuracy in tests equal to 0.952, and this proves the model’s sensitivity to identifying OA, characterized by a high true positive rate. But it was observed that specificity of the model was reduced in CNN and other models at some of the thresholds suggesting a general bias toward false positive. This raises the question for further developments to increase the specificity or general sensitivity for healthy joints as opposed to affected joints. According to the discoveries, deep learning, especially CNN, may help clinicians diagnose OA accurately, and with even higher speed as compared to the existing diagnostic system. In light of these findings, current models to improve the diagnostic tools to optimise the evaluation of OA should be fine-tuned.

Ähnliche Arbeiten