Generation of multidisease fundus photographs with code-free platform

Purpose To generate fundus photographs of multiple kinds of retinal disease, bypassing the requirement of coding technique. Methods The dataset contained fundus photographs of 10 categories of retinal conditions, with 500 fundus photographs in each category. We randomly divided the collected data into a training set (80%) and a test set (20%). Google Colaboratory was used to implement Pix2Pix to generate fundus photographs for each category. We compared the diagnostic abilities of ophthalmologists on both real and synthetic images. The diagnostic performance of the classification models trained on real, synthetic and combined data sets was also compared. Furthermore, the real and synthesised images were distinguished by ophthalmologists and artificial intelligence (AI) image detection websites. Results Fundus photographs of 10 categories were successfully synthesised using our method. The synthetic images showed slightly higher diagnostic accuracy by the three ophthalmologists than the real images (99.7% vs 98.7%, 98.0% vs 96.0% and 99.7% vs 94.3%; p=0.109). Training ResNet-50 and VGG-19 models with a combination of real and synthetic images resulted in significant improvements in accuracy, achieving 93.7% and 89.3%, respectively. Five residents achieved at least 92.5% accuracy in discriminating between real and synthetic images. In contrast, three AI image detection websites showed limited capability in this task, with a maximum accuracy of 51.2%. Conclusion Pix2Pix on Google Colaboratory can efficiently produce a diverse range of fundus images with typical characters.