Abstract
Purpose :
Image quality is a key challenge in retinal image analysis, with 20-45% of images typically being deemed unanalysable and discarded. These exclusions lower statistical power and introduce selection bias, but existing quality taxonomies are not fine-grained and might lead to unnecessary exclusions. The most commonly used taxonomy is the one underlying the EyeQ dataset, which provides a one-dimensional good-bad scoring. Automated methods for quality scoring like MCFNet, AutoMorph, or QuickQual are all based on the EyeQ taxonomy.
Methods :
We developed the Fundus Imaging – Nuanced Evaluation of Quality (FINEQual) taxonomy, which evaluates the visibility of macula, disc, and vessels separately on a four-level ordinal scale (Fig. 1) to provide more fine-grained information about image quality. 3 graders with relevant experience graded 100 images from UK Biobank. We compare the mean manual grades to the output of QuickQual for the probability of being bad (“p(bad”)). We chose QuickQual as it obtains state-of-the-art performance on the EyeQ dataset.
Results :
Fig. 2 shows the results of this assessment. The macula was scored as unusable in 21.7% of cases, the disc in 10.3% of cases and the vessels in 8.3% of cases. Graders agreed with mean Cohen’s weighted Kappa of 0.6176, 0.5586, and 0.6136 for macula, disc, and vessels respectively; which compares favourably with agreements described in the literature. QuickQual’s p(bad) score correlates with the mean grades for macula, disc, and vessels with Pearson correlations of 0.8793, 0.6089, 0.9063 and Spearman rank correlations of 0.8617, 0.5709, 0.8716 (all p<0.0001).
Conclusions :
We propose the FINEQual taxonomy which could provide more detailed information about image quality and thus allow for more targeted approaches to image quality exclusions. We observed good inter-grader agreement and found that existing one-dimensional scores correlate well with the visibility of macula and vessels, but poorly for visibility of the disc.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.