Fig 1.
Developed annotation tool for automatic radiographic scoring system.
Windows (a) are the image windows that show X-ray images to be annotated, and windows (b) to (d) are the annotation input windows. Since our tool works on the electronic medical record system, the annotators annotated the images during their spare time.
Fig 2.
Overview of preprocessing and training phase of our method.
Annotated images were resized and augmented before inputting to the DNN model. We trained two models to predict each label: subluxation and ankylosis. In the experiment, we also verified by replacing ResNet in the figure with several DNN models (AlexNet, DenseNet, ViT).
Table 1.
Results of subluxation classification.
Fig 3.
Average learning and ROC curves for subluxation classification.
Both results were obtained from the training of ResNet with Adam optimizer and Dataset 2.
Fig 4.
Examples of X-ray images in subluxation dataset.
The caption of each image indicates the prediction result.
Fig 5.
Average learning and ROC curves for ankylosis classification.
Both results were obtained from training ResNet with Adam optimizer and Dataset 2.
Table 2.
Results of ankylosis classification.
Fig 6.
Examples of X-ray images in ankylosis dataset.
The caption of each image indicates the prediction result.
Fig 7.
Visualized parts contributing to the prediction result using Grad-CAM.
The contribution to the model increases along with the red, yellow, green, and blue scales.
Table 3.
Related previous works and our study.