Abstract
Purpose :
The visibility of retinal structures on fundus images is a key factor for diabetic retinopathy (DR) screening. This study aims to evaluate if normalization of fundus images for illumination variations improve the detection of DR related lesions by professional fundus image graders (Moorfields Reading Centre).
Methods :
Four-field fundus image sets obtained from one eye (randomized) of 150 patients with diabetes mellitus attending the Rotterdam Eye Hospital after 01/May/2014 were included. Each image set was normalized for local variations in illumination, yielding uniform luminance and increased contrast of retinal structures (Adal et. al. WBIR 2014). Each four-field image set was combined into a large mosaic of color, red-free and normalized images for comprehensive retinal examination (Fig. 1).
For each eye, mosaics of Color+Red-free (CR) and Normalized+Color (NC) were shown separately and independently to 5 expert graders for detection of DR lesions and to assign DR level (ICDR severity scale). The number of red DR lesions (microaneurysms and haemorrhages) and severity of DR were analysed for both the CR and the NC mosaics.
Results :
Altogether 300 (150 CR/150 NC) mosaics were annotated; Table 1 summarizes the detected red lesions. Graders detected significantly more lesions (on average 122%) on NC mosaics compared to CR (Wilcoxon signed-rank test; p=0.03). In addition, graders were more likely to assign a higher DR severity level to NC mosaics compared to RCs (Table 1; proportional odds mixed model; odds ratio=3.4, p<10^-16); however there was only moderate agreement between the graders (ICC(2,1) =0.51).
Conclusions :
Fundus image illumination normalization improves the visibility of retinal features and hence facilitates better detection of red DR lesions. Higher DR severity levels were assigned after normalization. This warrants further research as its validity and clinical consequences must be established before illumination normalization can be used in clinical practice.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.