A STUDY COMPARING THE FUNDUS IMAGE-GENERATING MODEL'S PERFORMANCE

Abstract

Artificial intelligence technologies have been used much more often in recent years for processing images in medical research, including the analysis of medical data. However, the existence of personal and medical information presents several challenges for deep learning studies in the medical industry. Research projects suffer significant delays and financial losses as a result. Consequently, in order to stimulate research on deep learning in the medical domain, a medical image creation study was carried out in this study to identify the features of lesions using aberrant medical pictures inside medical data. A total of 356 fundus photos were used in this investigation, and three lesions, comprising images that were normal, were identified utilizing the 'Ocular Disease Intelligent Recognition' open dataset. The Res U-Net deep learning model, which generates images like the real fundus image by adding Residual Blocks to the U-Net structure which produces current fundus images, was employed in this work to generate fundus images. Furthermore, in order to compare the effectiveness of the current Res U-Net model and U-Net model in this investigation, three picture similarity indicators and ophthalmologist verification were used to objectively analyze and assess the fundus images produced by each model. The comparative assessment findings demonstrated that Res U-Net outperformed traditional models in all picture similarity measures, with Fréchet inception distance (FID) demonstrating an 8-fold improvement in performance. The average area under the curve, also known as or AUC, for all four lesions after the fundus picture produced in this study was validated by an ophthalmologist is Res U-Net 0.7415, U-Net 0.7705, indicating that the image that was generated was more similar to the original image. Subsequent studies will look into lesion insertion and deletion as well as patient data generation models with fundus images.