Roplight: The Early Detection of Retinopathy of Prematurity and Plus Disease

Overview

Retinopathy of Prematurity (ROP), observed in premature infants with low birth weight, is a leading cause of blindness worldwide. ROP shows no external signs or symptoms. Normal blood vessels may stop growing and cause new abnormal blood vessels to grow to provide nourishment to retina, eventually leading to retinal detachment. Currently, the gold standard to detect ROP is through meticulous examination by an expert ophthalmologist. In under-developed areas, where there are few ophthalmologists, screening for ROP is difficult and is the leading cause of childhood blindness. The objective of my research is to build a software and hardware solution that can be used by non-physician staff to screen ROP, especially in the early stages. With the goal of making early screening of ROP more effective and accurate, I have developed an automated ROP detection methodology that uses a combination of advanced image processing and machine learning techniques to detect and classify various zones, stages, and types of ROP. The proposed approach is an improvement over traditional algorithms that use retinal fundus images with handcrafted biomarkers. It offers a way to objectively measure the biomarkers for ROP by accurately segmenting blood vessels using ResUNet CNN architecture by classifying the ridge in Stage I and II using Mask R-CNN architecture and by measuring tortuosity and dilation indexes of blood vessels in severe forms of ROP, namely the Plus disease. The characterization, classification, and vessel segmentation methods proposed here are novel, clinically more accurate, and traceable. Overall, this automated model provides a complete picture of ROP Zone, Type, and Stage that trained medical staff can use to diagnose ROP or refer to an expert ophthalmologist for further examination. I have also developed a cheaper imaging technology by designing a smart phone lens mount to capture high quality photographs of the eye. The fundus image capture through the proposed prototype consisting of 3D adaptor, smartphone, and 20 Diopter(D)/28D condensing lens provides a low-cost alternative to the expensive RetCam. By automating the detection of ROP in early stages, the proposed approach will complement efforts in telemedicine and traditional methodologies to efficiently and accurately diagnose preterm babies who are at risk of ROP, on time and in underserved areas.