Multicenter comparison of deep learning models for Kellgren-Lawrence grading of knee osteoarthritis

In a multicenter study of 1550 patients, VGG-16 achieved AUC 0.93 (accuracy 0.73) for Kellgren-Lawrence classification of knee OA on AP radiographs, but external validation accuracy fell to 0.65. Adding lateral radiographs or clinical data (age, gender, BMI) did not improve perf…

• VGG-16 significantly outperformed ResNet-50, DenseNet-121, Inception_v3, and EfficientNetb0 (p<0.001) with internal accuracy 0.73, sensitivity 0.72, specificity 0.93, and AUC 0.93.
• External validation on two centers showed accuracy 0.65 and AUC 0.90.
• Lateral radiographs and patient clinical data (age, gender, BMI) did not significantly improve KL classification performance.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.