Spatial distribution modelling of retinal lesions with application to retinal capillary non-perfusion in malarial retinopathy.

Abstract

Purpose : Evaluation of retinal disease typically focuses on assessing the number or size of lesions, but this ignores the fact that the retina is made up of diverse anatomical regions. Adding spatial location to analyses may increases the power to detect retinal disease features. We use capillary non-perfusion (CNP) as an example because CNP is common and relevant for diabetic retinopathy, retinal vein occlusion and posterior uveitis. We use malarial retinopathy as an example with death as a clear endpoint.

Methods : Comatose Malawian children with cerebral malaria and malarial retinopathy were recruited into a prospective observational study and underwent dilated ophthalmoscopic exam and fluorescein angiography (FA) on admission. We divided the macula and mid periphery (6.3 mm radius) into 48 sectors (see Figure) and used percentage of CNP pixels in each sector to characterise the spatial distribution of lesions across the retina. We developed a hierarchical statistical model to represent the CNP spatial distribution with goniometric functions, estimated the model parameters by maximum likelihood and tested associations with the primary outcome, death.

Results : 172 cases were recruited (2009 to 2014) amongst whom 138 had fair, good or excellent quality FA images (120 survived and 18 died). Using our spatial model we found that in malarial retinopathy macular CNP generally occurs around the fovea and in the temporal macula with increased CNP in death (p=0.03). The highest CNP differences were in the macular temporal area (27% vs 31%, for survived vs died, p=0.01) and peripheral temporal area (15% vs 22%, for survived vs died, p<0.01). Less strong associations were found using two-sample t-tests, with insignificant associations between the macula and outcome (all 24 p-values >0.05) and one significant sector for periphery (p-value=0.02, other 23 p-values >0.05). Lower image quality led to higher within-eye variation (17% and 28% higher SD for good and fair vs. excellent quality images).

Conclusions : We developed a novel statistical modelling approach to analyse spatial lesion in the retina. We illustrated the spatial approach in the analysis of lesions of macular CNP and showed that it provides both a richer description of lesions than existing methods and increased power to detect associations with clinical outcome. This technique is relevant to a range of retinal diseases.

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.