Augmenting Fundus Images using Color Variants Based on HSV Color-space Quartile Values

Fundus coloration is caused by the retinal pigment epithelium containing melanin and choroidal blood vessels. Retinal pigmentation is influenced by ethnicity, age, refraction, skin, hair, and iris pigmentation. Four color variants were suggested by ophthalmologists observing public fundus datasets. When training on datasets with limited color variants, the universality and accuracy of neural network models could be affected. To address this problem, we created a specific and novel fundus augmentation method using exact fundus color variants. This augmentation is domain-specific, as it incorporates biological and clinical knowledge about fundus color variation. The DBSCAN clustering method was used to divide healthy images from three public datasets into seven different fundus color variants according to the HSV color-space quartile upper limits. Based on these values in each cluster, transformations of the original images were performed. Consequently, seven new biologically meaningful augmentations per image were created to enrich the dataset for neural networks training.