A Study on the Performance Comparison of the Fundus Image Generation Model

Abstract

Recently, in the medical field, research using medical data based on artificial intelligence image processing technology has been actively conducted. However, research in the medical field using deep learning has many difficulties in obtaining data because personal information and medical information exist. This causes a lot of time and economic losses in conducting research. Therefore, in this study, a medical image generation study was conducted to realize the characteristics of lesions based on abnormal medical images among medical data to activate deep learning research in the medical field. In this study, three lesions, including normal images, were screened using the 'Ocular Disease Intelligent Recognition' open dataset, and a total of 356 fundus images were used. The deep learning model used for generating fundus images in this study is Res U-Net, which adds Residual Block to the U-Net structure that generates existing fundus images, and produces an image similar to the actual fundus image. In addition, for the performance comparison between the existing U-Net model and the Res U-Net model in this study, the fundus images generated by each model were quantitatively compared and evaluated through three image similarity indicators and ophthalmologist verification. Comparative evaluation results showed that Res U-Net showed higher results in all image similarity indicators than conventional models, especially Fréchet inception distance (FID) showed 8 times better performance. As a result of the ophthalmologist's verification of the generated fundus image, the average area under the curve (AUC) of the four lesions was U-Net 0.7705, Res U-Net 0.7415, indicating that the image generated in this study was more similar to the original image. As a future study, the addition and removal of lesions and patient information generation models will be studied based on fundus images.