Convolutional Neural Network Performs Similarly to Orthopedic Surgeons in Radiographic Grading of Knee Osteoarthritis.
TL;DR
CNN-based grading demonstrated performance comparable to orthopedic surgeons, supporting its potential role as a screening or triage tool to improve consistency and efficiency in clinical workflows.
Abstract
Background
Convolutional neural networks (CNNs) have shown promise in automated grading of knee osteoarthritis (OA) using the Kellgren-Lawrence (KL) scale. This study aimed to externally validate a CNN model against KL grading by fellowship-trained orthopedic surgeons.
Materials And Methods
CNN architectures (VGG16, ResNet34, DenseNet196, EfficientNetV2) were trained on 8,260 knee radiographs. Model performance was evaluated using accuracy, area under the curve (AUC) on a receiver operating curve, and F1 score. External validation was performed by comparing model outputs to KL grades assigned by 20 orthopedic surgeons using a blinded survey of 10 radiographs. Agreement was assessed using intra-class correlation coefficients (ICC) and Bland-Altman analysis.
Results
EfficientNetV2 demonstrated the highest performance (AUC: 0.83; accuracy: 71%). The model achieved a mean absolute error of 0.58 KL grades relative to the average physician score and performed best at identifying KL0 and KL4 cases, with reduced performance for intermediate grades, particularly KL1. Average physician and model KL scores did not differ significantly (P = .23). Agreement between physicians and the model was high for both good-performing (ICC 0.93) and poor-performing images (ICC 0.94), exceeding inter-physician agreement (ICC 0.74 and 0.72, respectively). Bland-Altman analysis demonstrated minimal bias (-0.25 KL grades), indicating that physicians graded more conservatively than did the model.
Conclusion
CNN-based grading demonstrated performance comparable to orthopedic surgeons, supporting its potential role as a screening or triage tool to improve consistency and efficiency in clinical workflows.