Table 1.
Summary of dataset composition used in each experiment.
ODCR stands for Ophthalmologist Diagnosis Consensus Rate. For the test dataset, we selected 100 normal eye images and 96 strabismus eye images from the data collected initially, which had an ODCR of 80% or higher. The generated data was exclusively used for training and not used for validation or testing.
Fig 1.
Test results (Accuracy(a) and AUC(b)) for 6 experiments. L1 shows the results without data augmentation and generated data addition. L2 represents the results with data augmentation only. L3 to L6 demonstrate the test results after adding 500, 1000, 2000, and 3000 generated data to the training dataset, followed by data augmentation.
Fig 2.
Visualize overfitting problem: Difference between training and validation loss.
Visualization of the effectiveness in addressing overfitting problems when performing classification tasks based on the ResNet50 and ResNext101. Each experiment shows the difference in accuracy and loss between training and validation at the end of training. (a), (c) represents the accuracy difference, and (b), (d) represents the loss difference. (a) and (b) represent the results of the ResNet50 model, while (c) and (d) represent the results of the ResNext101 model.
Fig 3.
Data samples generated based on StyleGAN2-ADA. Both squint eye and normal eye samples were extracted from the Generator of the trained model, which was trained with an ODCR of 80% as the criterion. (a) represents the generated strabismus eye sample, and (b) represents the generated normal eye sample.
Table 2.
Generation performance according to ODCR.
Fig 4.
Visualization of the activation regions using GradCAM for the classifier(ResNet50).