Training U-Net with Proportional Image Division for Retinal Structure Segmentation

Cup and optic disc segmentation has become one of the main objects of study in the field of creating and improving machine learning-oriented models due to the importance of vision for human beings and the ability to assist physicians in diagnosing ocular problems. Within this context, this study presents a new method based on the proportional division of images concerning features extracted from the sample set. These samples go through a pre-processing step involving image resizing before going to deep feature extraction and K-means clustering, thus dividing the set for validation and training. Soon after, the amount of samples is increased through data augmentation before going on to the U-Net training. The proposed method has been evaluated on the public RIM-ONE and DRISHTI-GS datasets, and presented promising results in the segmentation of both structures, with emphasis on obtaining the value of 92.2% of Dice for the segmentation of the optic cup in the DRISHTI-GS test dataset and 95.9% of Dice for the optic disc in the RIM-ONE.