Abstract
Background
Accurate assessment of lateral knee radiographs is crucial for evaluating knee biomechanics and guiding treatment decisions. However, manual evaluation is time-consuming and subject to variability. This study aims to develop a machine learning model that can automatically classify lateral knee X-rays, paving the way for automated measurement of important biomechanical parameters like posterior tibial slope and patella tendon indices.
Methods
929 random lateral knee X-rays, with Kellgren Lawrence (KL) grade 0 and 1 were extracted from the Osteoarthritis Initiatives (OAI) publicly accessible database. We randomly split the 929 images into 729 images for the training set and 200 images for the test set. The images were evaluated for quality and classified into three categories: 'Excellent', 'Good', and 'Bad'. Region of interest was identified and cropped using a deep learning object detector. The images were resized to 320 * 320 and augmented. We utilized fine-tuning versions of Convolutional Neural Networks (CNN) architectures, with subsequent 5-fold cross validation to help with hyperparameter tuning. Model performance was evaluated with area under the receiver operating characteristic curve (AUC) and Accuracy.
Results
ResNet was the most accurate model, with a composite AUC of 0.979 (CI= [0.964–0.99]). The highest accuracy was achieved correctly classifying the ‘Bad’ class. The confusion matrix showed that classifying 'Bad' and 'Excellent' classes was simpler than 'Good,' consistent with expert human annotators. Saliency maps highlighting the most predictive area of the X-ray exhibit a focus on alignment of the posterior femoral condyle.
Conclusion
We successfully developed and validated a high-performing deep learning model for classifying lateral knee X-ray images. Ultimately, the ability to extract accurate biomechanical data from routine lateral radiographs through automated analysis has potential to revolutionize orthopedic care delivery, enabling cost-effective assessments, aiding surgical planning, and facilitating outcome evaluation – paving the way for improved diagnostic capabilities and better patient outcomes.