Distinguishing image quality from gradeability: the relationship between quality and gradeability for color fundus photographs in glaucoma detection

Abstract

Purpose :
Most imaging tools employ quality metrics to assess success or failure in image capture. While quality is a useful metric, it often serves as a proxy for gradeability - whether a fundus image can be graded for a given ophthalmic condition. This work aims to investigate the relationship between gradeability and quality and to explore clinically impactful differences. It compares a deep learning (DL) gradeability score to other image quality metrics for photos with various types and severities of imaging issues.

Methods :
Clinically relevant global (blur, increased and decreased brightness, contrast, noise) and local (halo, hole, spot, cropping) degradations were applied to 100 photos from the Ocular Hypertension Treatment Study (OHTS) at varying levels of severity, resulting in 120,000 total images with simulated degradations(Fig. 1). Gradeability (using a previously published DL model) and quality (using both standard and DL-based approaches) were used to assign gradeability and quality scores, respectively, to all images for a total 1.08 million scores. The relationships between quality, gradeability, and glaucoma severity, as measured by visual field MD, were then evaluated for each type of degradation using R2.

Results :
Quality and gradeability were highly correlated for most forms of clinically relevant degradations (Fig. 2). However, cropping, which had a large impact on quality, had little effect on gradeability until the extent of crop obscured the optic nerve head. As such, R2 for gradeability and quality for all single-sided crops were low for all quality models, with the exception of a transformer-based quality model (clipiqa+vitL14+512). Also, while there was no relationship between quality and MD Index (R2=0), these were related for gradeability (R2=0.21, p<0.0001).

Conclusions :
Gradeability is not equivalent to quality. Glaucoma gradeability, the ability for a fundus photo to be graded for glaucoma, is a more direct way of determining the utility of captured images for glaucoma clinical and research purposes. Glaucoma severity correlates with gradeability, suggesting that the more significant or more obvious the disease, the more the image may remain gradeable even in the presence of imaging issues. A paradigm shift from use of quality to gradeability to assess value of captured images may provide clinical and research utility.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.