Keywords:
Bones, Extremities, Conventional radiography, Digital radiography, Neural networks, Computer Applications-General, Education, Education and training
Authors:
A. Matsushima1, C. Tai-Been2, T. Okamoto1, S.-Y. Hsu2, J. Ryu-I2, N. Itayama1, T. Ishibashi1, K. Fukuda1; 1Tokyo/JP, 2Kaohsiung City/TW
DOI:
10.26044/ecr2021/C-10411
Methods and materials
1. Definition of terms
Images that do not require re-exposure are termed “Pass”, and images that require re-exposure are termed “NG”.
2. Parameter settings
For each of four models (AlexNet, VGG16, ResNet50, and ResNet101), we changed the batch size and input size parameters, executed the program, and verified the accuracy.
In addition, we also modified only the size of each layer without changing the number of layers, and used it with minimal modifications to the original model. The settings described above were verified and the accuracy was calculated.
3. Image database
The training images were lateral views of the knee joint of a lower-limb phantom taken by a digital radiography device. The images were obtained in DICOM format and then exported to jpg format. Two types of image databases were considered: one containing images at various resolutions, and the other containing images at the same resolution. Figure 3 shows the details of the images in the databases.
4. Evaluation criteria
The current imaging criteria of Teikyo University Hospital, Japan, were used for evaluation of the lateral views of the knee joint. As shown in Fig. 4, images with a distance of >=5mm between the medial and lateral condyles were designated “NG”, and those with a distance of <5mm were designated “Pass”.
5. Image classification methods
- Using the model with the highest accuracy from the results of 2, the images were classified as, NG or Pass, based on the exclusion criteria, and the accuracy of each are shown.
- The model with the highest accuracy from the results of 2 was used for five-class image classification, as shown in Fig. 5. NG was further refined to “abduction from evaluation criteria”, “adduction from evaluation criteria”, “internal rotation from evaluation criteria” and “external rotation from evaluation criteria”.
6. Testing on new images
Based on the training results of the CNN model with the highest accuracy, 99 new images were evaluated and the accuracy of the model was calculated. These new images were X-ray images of a phantom that were taken during class experiment by a second-year student in the Faculty of Medical Technology.
7. Development environment
The development environment used is shown in Fig. 6.