1. Two-class classification
1.1 Results of the batch size study for Image Database No.1
The batch size study was performed ten times for each trial, and the results for image database No.1 are shown in Figs 7-11. The highest accuracy was obtained for ResNet101 with a batch size of 16.
1.2 Results of the batch size study for Image Database No.2
The batch size study was performed ten times for each trial, and the results for image database No. 2 are shown in Figs 12-16. The highest accuracy was obtained for ResNet101 with a batch size 6.
1.3 Results of the input image size study for Image Database No.2
The batch size was then fixed based on the results obtained in the above batch-size verification, using ResNet101. We modified part of the model to resize the input data. The results are shown in Fig. 17. The highest average accuracy and smallest standard deviation was obtained with an input size of 256. There was no significant difference in t-test statistical analysis among input sizes of 256, 384, and 512. We decided to use ResNet101 with a batch size of 6 and an input size of 256.
2. Five-class classification
One of the problems with adapting a CNN to images is overfitting. Overfitting refers to the phenomenon when a network learns a function with very high variance in order to model the training data perfectly. Data augmentation encompasses a suite of techniques that enlarge the size of a training dataset. In the present study, two of the primary methods for geometric transformation augmentations were used to increase the number of images.
Figure 18 shows the results of implementing data augmentation, increasing the number of training images, and adapting ResNet101, with a batch size of 6 and an input size of 256. The images were classified into five-classes: “Pass”, “abduction”, “adduction”, “internal rotation”, and “external rotation” based on the evaluation criteria. The highest accuracy was achieved for ~4000 images.
3. Two-class classification test
Figure 19 shows the results of evaluating 99 images using the two-class classification model with the highest accuracy. This accuracy of the model was 0.9293.
4. Five-class classification test
Based on the training results for the CNN model with the highest accuracy, 99 new images were classified into five classes and the accuracy was calculated. As shown in Fig. 20, the accuracy was 0.9293.