Abstract: BACKGROUND: Knee osteoarthritis is a common degenerative disease that significantly impacts patients' quality of life and increases the societal healthcare burden. Early and accurate diagnosis of knee osteoarthritis is crucial for the treatment and prognosis of patients. Traditional diagnostic methods are not only subjective and time-consuming but also do not guarantee consistently high accuracy.
OBJECTIVE: To develop an automatic diagnostic method for knee osteoarthritis based on deep learning, utilizing deep learning networks to improve diagnostic accuracy and efficiency.
METHODS: A new network model, YOLOV8-ViT, was proposed by replacing the backbone network of YOLOv8n with the Efficient-ViT network and incorporating attention mechanisms for the automatic identification and classification of X-ray images of knee osteoarthritis. The experimental dataset included 5 078 X-ray images of patients with knee osteoarthritis obtained from the Third Affiliated Hospital of Guangzhou University of Chinese Medicine. Three imaging physicians annotated the sites of knee osteoarthritis and classified them according to the Kellgren-Lawrence grading standard using Labelme software, and the results were combined. The evaluation indicators used in this study included Precision, F1 score, mean average precision (mAP), Recall, val/box_loss, val/cls_loss, and val/dfl_loss.
RESULTS AND CONCLUSION: The experimental results showed that the YOLOV8-ViT model outperformed the YOLOv5n, YOLOv8n, and YOLOv9n models in terms of precision, mAP50, mAP50-95, F1 score, and Recall, while lowering val/box_loss, val/cls_loss, and val/dfl_loss by 0.496, 0.45, and 0.523; 1.037, 0.305, and 0.728; and 0.267, 0.654, and 0.854, respectively. These experimental data validate that this model has high detection accuracy.

Key words: knee osteoarthritis, deep learning, YOLOv8, Transformer, object detection, detection precision