Fig 1.
The study workflow.
Fig 2.
Kellgren-Lawrence grade diagram (The left image depicts a normal knee joint’s X-ray image, and the right one illustrates an X-ray from a patient with KOA).
Fig 3.
In this study, we employed Depthwise Separable Convolution to significantly reduce the model's parameters, decreasing complexity, thus expediting training and reducing the risk of overfitting. This technique enabled us to understand the images from various angles and scales, extracting a plethora of osteoarthritis features. Additionally, we incorporated Residual Blocks, effectively alleviating the vanishing and exploding gradient issues common in deep networks. This adaptation allowed for more profound network training. Leveraging pre-trained DenseNet, we extracted valuable information such as edge, color, texture, and other features from osteoarthritis images, significantly enhancing the efficiency and accuracy of our model. Moreover, we fused feature extraction from multiple layers, markedly improving the model's sensitivity compared to single models. This enhancement led to superior predictive accuracy.
Fig 4.
Result of five-fold cross-validation.
(a) Training process of OHC model. (b) ROC curve of OHC model. (c)-(h) ROC curve of DenseNet, EfficientNet, MobileNet, NASNet, ResNet, and ShuffleNet, respectively (the black middle line represents the best two models).
Table 1.
Model performance in five-fold cross-validation.
Table 2.
Model performance in external validation.
Fig 5.
Results of external validation.
(a) Confusion matrix of the OHC model. (b) ROC curve of OHC model. (c)-(h) ROC curve of DenseNet, EfficientNet, MobileNet, NASNet, ResNet, and ShuffleNet, respectively.
Fig 6.
The heat map of different categories (a) the heat map of healthy knee joint.
(b) the heat map of KOA patient's knee joint.