Improved Segmentation of Infant Retinal Images and Quantitative Vascular Analysis

Retinopathy of Prematurity is a leading cause of visual impairment in preterm infants and is characterized by dilation and tortuosity of the retinal blood vessels in the plus disease stage. However, the diagnosis of plus disease is subjective and qualitative; hence, quantitative methods and computer-based image analysis are required to improve the objectivity of the diagnosis. In this study, we proposed a computer-based image analysis method aimed at segmenting blood vessels and the optic disc in retinal images and providing quantitative features of the vessels to assist doctors in diagnosing plus disease. This method comprises two main stages. In the first stage, we used fundus images of a preterm infant with manually annotated vessel segmentation labels to train U-Net3 network, which is a U-Net network with the dual attention modules. Simultaneously, we used images with optic disc segmentation labels to train U-Net1 network, which is a U-Net network with half of the channels. The F1 score of the vessel segmentation network was 0.8116, and the sensitivity was 0.8273. The F1 score of the optic disc segmentation network was 0.9346, with a sensitivity of 0.9395. In the second stage, we calculated the vessel tortuosity in different retinal regions of the vessel segmentation images using least-squares linear regression. In addition, we computed the vessel density and width within the optimal region of interest, which is defined as a radius four times the diameter of the optic disc. The quantitative results revealed that the vessel tortuosity, density, and width obtained in this study can be used as diagnostic features.