Comparison of expert graders to computer-assisted image analysis of the retina in retinopathy of prematurity.

Abstract

To determine correlation of width and tortuosity between expert graders and computer-assisted image analysis of the retina in narrow-field images of eyes with retinopathy of prematurity. 11 digital images were selected based on severity of retinopathy of prematurity (ROP). Narrow field images were analysed for width and tortuosity of vessels using computer-aided image analysis of the retina (CAIAR), an image analysis software, and by four ROP experts. Spearman correlation coefficients (ρ) assessed the correlation of CAIAR grading with expert grading. Intra-class correlations assessed agreement among graders. Width and tortuosity were compared among severity of ROP and treatment status using analysis of variance and generalised estimating equations. Expert measurements correlated well with measures from CAIAR for venule width (ρ=0.57-0.66) and arteriole tortuosity (ρ=0.71-0.81). Measurements from four graders agreed moderately well (intra-class correlations were 0.49 and 0.69 for venule width and arteriole tortuosity, respectively). Increased severity of ROP (no pre-plus/plus, pre-plus, plus) was associated with larger width (linear trend p=0.02 in two graders) and tortuosity (linear trend p<0.03 in all graders). Tortuosity measurements by CAIAR and graders were statistically different between treated and untreated eyes (p<0.002). We found moderate correlation between expert graders' assessment of vessel tortuosity and width and CAIAR using narrow-field images.