Deep learning for retinopathy of prematurity screening

Retinopathy of prematurity (ROP) is a neurovascular disorder of retina, characterised by abnormal fibrovascular proliferation at the boundary of the vascularised and avascular peripheral retina. Globally, it is estimated that 19 million children are suffering from visual impairment.1 Of those, ROP accounts for 6%–18% childhood blindness,2 causing significant psychosocial impact on the child and the family.3 According to the Early Treatment for Retinopathy of Prematurity (ETROP) trial,4 early treatment has shown to be beneficial to improve the visual acuity of the high-risk patients with ROP, although 9% still eventually became blind. Thus, early screening with regular monitoring is extremely crucial. The at-risk groups are babies who are born preterm or those with neonatal morbidity, for example, respiratory distress syndrome, infection and hyperglycaemia.5 These groups of neonates usually require high oxygen demand due to the systemic issues. Oxygen regulation is important for normal retinal vascular development. In the UK, the current guidelines recommend that babies born at <32 weeks or with birth weight <1.5 kg should be screened for ROP.6 For the medically unstable infants who require high supplemental oxygen, the screening is recommended to be done earlier, although the screening criteria may vary slightly between different countries around the world. Based on the International Classification of ROP, ROP is divided into five stages (table 1) with or without plus diseases.7 Early recognition of plus diseases is extremely crucial for initiation of treatment. The disease involvement is usually documented in terms of location and the extent of clock hours. For location, it is divided into zones 1–3 (figure 1). It is important to detect the patients with prethreshold (type 1 ROP and type 2 ROP) and aggressive posterior ROP (also known as ‘rush disease’ previously). Serial diagnostic examinations should be performed until each eye …