Abstract
Purpose :
One of the main limiting factors of classification algorithms in the medical imaging domain is the lack of relevant class-specific data, especially for rare disease conditions, which often leads to suboptimal performance of classification models. In this study, we experimented the use of a Data Efficient Generative technique to synthesize class specific DR images. Generated images were used to improve the diversity of input datasets for the training of DR classification models, and the performance of these models were evaluated against the performance of models trained on existing DR datasets.
Methods :
StyleGAN2-ADA was used as the generative technique to synthesize class-specific diabetic retinopathy images. The dataset consists of 5 classes: No DR (0), Mild DR (1), Moderate DR (2), Severe DR (3) and Proliferative DR (4). The study has 3 stages: 1) Balance out the EyePacs DR dataset with oversampling and use it to train a Densenet-121 classification model. 2) Train StyleGAN2-ADA to synthesize class specific DR images using conditional class generation. 3) Train and evaluate another Densenet-121 classification model with the original dataset supplemented by GAN-generated images while keeping the testing dataset and hyper parameters constant.
Results :
The training of the StyleGAN2-ADA model converged to a Frechet Inception Distance (FID) of 3.465. The accuracy of the Densenet-121 model trained on the original balanced Kaggle dataset was 0.60. This increased to 0.76 when the dataset was supplemented with GAN-generated images. The f1-score for class 0 changed the least, and the f1-scores of other classes increased to match the performance of class 0. This indicates that the poor performance was originally the result of oversampling for classes with deficient data and that performance improves when each class has an equally diverse set of images.
Conclusions :
The performance of models trained on DR datasets supplemented with GAN-generated images perform significantly better on classification tasks. This study validates the potential for StyleGAN2-ADA to generate realistic DR images and highlights the use of generative models as a data supplementation techniques as compared to conventional oversampling methods.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.