Quality assessment of non-mydriatic fundus photographs for glaucoma screening in primary healthcare centres: a real-world study

Introduction

As the world’s leading irreversible and age-related blinding eye disease, glaucoma is often asymptomatic in its early stages,1 making population screening crucial for early detection. A systemic review estimates that the number of patients with glaucoma worldwide will increase to 111.8 million in 2040.2 Along with an ageing population, improved life expectancy, changing environments and behavioural patterns, healthcare systems are facing unprecedented challenges in meeting the current and anticipated demands for eye care. Although population-based screening of angle-closure glaucoma and open-angle glaucoma has been shown to be cost-effective in China,3 it is still not a routine screening programme owing to ophthalmic personnel shortages and financial constraints. Encouragingly, telemedicine and artificial intelligence (AI) algorithms have revolutionised the field of eye health screening. Telemedicine enables ophthalmologists to evaluate and deliver care remotely, which facilitates a more efficient and equitable distribution of limited healthcare resources.4–6 AI-assisted screening using colour fundus photographs (CFPs) has demonstrated high sensitivity and specificity.7–12 Additionally, deep learning (DL) systems have made it possible to predict the onset and progression of glaucoma.11 12 Recently, Wang et al 13 developed a Markov model and concluded that annual AI telemedicine screening was the most cost-effective strategy in China.

Glaucoma screening is highly dependent on various imaging devices. Among them, fundus photography has become a valuable tool due to its high feasibility, important diagnostic value and wide availability. Obtaining high-quality CFPs is crucial for the development and application of DL systems, especially for the internal set used for training and validation.7–12 However, fundus photographs acquired during real-world glaucoma screening are often non-mydriatic with short shooting times and can be obscured due to media opacities such as cataracts. A certain proportion of fundus photographs obtained in population-based glaucoma screening may be deemed ‘unreadable’. Poor-quality images limit the application of AI in the healthcare industry, yet little attention has been given to the image quality of real-world fundus photographs.

In light of this, we employ a new evaluation criterion to assess the quality distribution of NMFPs obtained during real-world glaucoma screening and to analyse the factors that impact image quality. The results can serve as a valuable reference for enhancing the quality of fundus photographs in primary healthcare centres, optimising the screening approach for eye diseases and improving the efficiency of screening for those conditions.

Methods

This cross-sectional study was part of a randomised controlled trial (RCT) of community glaucoma screening in Yinzhou District. Informed consent was obtained from all participants. The protocol has been published on the website (http://www.medresman.org.cn/) with registration number ChiCTR2200059277.

Quality grading for fundus photographs

This study established a criterion for grading fundus images to facilitate rapid population screening of glaucoma, which is mainly based on the photographic location and clarity. If the optic disc was not fully located or clearly reviewed by the reading staff, the image quality was considered unreadable. If the optic disc was clearly visible but the macular area and blood vessels were blurry, the image quality was considered readable with only the optic disc visible. If the optic disc and macular area were clearly visible but flaws such as artefacts or defocus existed, the image quality was considered readable with both the optic disc and the macula visible. If the optic disc, macular area and blood vessels were clearly visualised without obvious flaws, the image quality was considered reliable. Therefore, the image quality was divided into four levels, and details are shown as follows:

The image quality was deemed reliable according to the established criterion17: the photographs were free from defocus, blurs, underexposure or overexposure, and both the macula and the optic disc were visible and clear, as demonstrated in figure 1A.

Request permission

Figure 1

Quality grading criterion for non-mydriatic fundus photographs. (A) Reliability fundus photographs. (B) Readable (both optic disc and macula visible): the photograph was mild defocus, but the results could still be interpreted. (C) Readable (both optic disc and macula visible): artefacts were discernible that did not affect the reading results. (D) Readable (optic disc only visible): the photograph was slightly defocused, but the results could still be interpreted. (E) Readable (optic disc only visible): artefacts are discernible that did not affect the reading results. (F) Unreadable: the photograph was seriously defocused that made it impossible to interpret the results. (G) Unreadable: large artefacts existed that affected the reading results. (G) Unreadable: the images were under exposed, so the optic disc was not visible.

The image quality was considered readable with both the optic disc and the macula visible. Although there was mild defocus in the photographs (as shown in figure 1B), the results could still be interpreted. Furthermore, the photographs contained easily discernible artefacts (figure 1C) that did not affect the reading results. Both the macula and the optic disc were visible in the photographs.

The image quality was considered readable with only the optic disc visible. Although the photographs were slightly defocused (as shown in figure 1D), the results could still be interpreted. There were easily recognisable artefacts (figure 1E) that did not affect the reading results. Only the optic disc and its supratemporal and infratemporal regions were visible in the fundus photographs.

The image quality was deemed unreadable in photographs that were seriously defocused (figure 1F) and contained large artefacts that made it impossible to interpret the results (figure 1G). The images were also underexposed, and the optic disc was not visible (figure 1H).

Among them, readability included images that were readable with only the optic disc visible, images that were readable with both the optic disc and the macula visible, and reliable images. Unreadability merely included unreadable images. Reliability included reliable images. Unreliability included unreadable images, images that were readable with only the optic disc visible, and images that were readable with both the optic disc and the macula visible.

Results

In total, 17 232 images from 8616 subjects (3545 male and 5071 female) were analysed (table 1). The median age of the subjects was 68 years (IQR 64~73 years). The median logMAR PVA of both eyes was 0.22 (IQR 0.10~0.40), and the spherical equivalent (SE) of both eyes was +0.5 diopters (D) (IQR −0.38~+1.13D). The median IOPs of the right and left eyes were 12.9 mm Hg (IQR 10.9~15.0 mm Hg) and 13.3 mm Hg (IQR 11.2~15.5 mm Hg), respectively. The IOP difference between bilateral eyes was statistically significant (p<0.001). Media opacity was observed in 5330 (61.9%) right eyes and 5563 (64.6%) left eyes, and the difference was statistically significant (p<0.001). The difference in the astigmatism axis between bilateral eyes was also statistically significant (p<0.001).

Discussion

This study first reported the distribution of image quality for glaucoma screening and related influencing factors in China. The fundus images of previous studies mostly originated from the clinic and mydriatic,21 22 and image quality was graded for the detection of disease-specific features. However, fundus abnormalities in large-scale glaucoma screening in communities vary widely. Because this study was conducted in the context of comprehensive eye health screening primarily focused on glaucoma, image readability was the primary concern. Accurate diagnosis of glaucoma relied on high-quality fundus photography, with particular emphasis on capturing the discrimination range of the optic disc and its surrounding two papillary diameters. This was critical for identifying the vertical cup-disc ratio, as well as the width of the superior temporal or inferior temporal disc edge and retinal nerve fibre layer.23 The readability of bilateral image quality was found to be above 90%, which was a higher rate compared with previous telemedicine programmes that used NMFP.24–26 This difference could be due to variations in the definition of unreadable images. Additionally, multiple age-related eye diseases tend to coexist and exacerbate each other in elderly individuals.27 The primary objective of screening was to refer individuals with suspected eye diseases and prevent cases of missed diagnosis. Comprehensive eye health screening could serve as a relatively cost-effective screening model, which would help to address issues related to high screening costs and low efficiency. As a result, it was recommended that four specific areas of a single-field fundus photograph—the optic disc, macular area and the first branch of the upper and lower retinal vascular arch—should be clearly visible as the standard for determining image reliability.28 However, the reliability of bilateral image quality in our study was found to be less than 12%, which was lower than previously reported.29 This indicated that the quality of real-world NMFPs fell far short of meeting this criterion. During the screening process, factors such as improper eye fixation, miosis, blepharoptosis, inaccurate focus, excessive or insufficient exposure, and lens opacity might contribute to a decline in the quality of fundus photographs. This could, in turn, distinctly impact the overall effectiveness of NMFP as a tool for population-based screening of fundus diseases in real-world settings.

Our study showed that the image readability and reliability of the right eyes were consistently better than those of the left eyes. There could be several reasons for this, including the photography sequence and insufficient interphotograph intervals, which could lead to inadequate pupil dilation and the proportion difference of media opacity, subsequently, more unreadable or unreliable images for the left eyes. In addition, the postillumination pupil response (PIPR)30 31 was sustained pupil constriction that occurred after light cessation, of which redilation velocity was its quantified metric. The PIPR redilation velocity has been found to increase with decreasing irradiance, shorter stimulus duration and longer wavelengths.32 A previous study30 reported that the maximum duration of PIPR in healthy participants was 83.0±48.0 s for 1 s pulsed light stimulation. Lamirel et al 28 found that taking high-quality photographs with an interphotograph interval of 30–90 s was 3.8 times more effective than intervals of <15 s. Our glaucoma screening was integrated into the community health examination, with a maximum of nearly 130 subjects being examined in the morning. We speculated that the bilateral photography interval during onsite screening fell within the PIPR period, especially for the left eyes, which worsened the image quality. These findings were essential for community-based screening programmes that primarily relied on NMFPs. A prospective study might be conducted in the future to find a suitable photographic interval.

This study also showed that older age, media opacity and poorer VA were associated with a decreased likelihood of obtaining readable fundus images, and older age was also associated with significantly decreased odds of image reliability. These findings were consistent with previous studies based on NMFPs.24 28 29 33 It was likely that older participants were more vulnerable to ocular media opacity and miosis, which could deteriorate image quality. Previous studies34–36 found that unoperated cataracts were a primary cause of visual impairment and blindness. Hark et al 24 established that the rate of visually significant cataracts in participants with unreadable images was twice as high as that in readable images (p<0.05). Therefore, the image readability of subjects with poorer VA was susceptible to media opacity.

We observed a novel finding that a higher absolute value of SE was associated with lower odds of fundus image readability and reliability. This could be attributed to poor fixation in individuals with high myopia or hyperopia, which might limit the fundus camera’s ability to fully compensate for ametropia. Our findings also indicated that oblique astigmatism was associated with a lower likelihood of image readability compared with with-the-rule (WTR) astigmatism, and against-the-rule (ATR) astigmatism was associated with a significantly higher likelihood of image reliability compared with WTR astigmatism. Regular astigmatism takes three different forms,32 and previous studies have shown that corneal irregular astigmatism was greatest in eyes with oblique astigmatism, even after adjustment for age in a stepwise multiple regression model. This was followed by WTR astigmatism and then ATR astigmatism.32 Corneal irregular astigmatism is known to increase with age and various corneal diseases.37–39 We speculated that eyes with oblique astigmatism have the largest corneal irregularity astigmatism, followed by WTR astigmatism and then ATR astigmatism, so the image readability of eyes with oblique astigmatism was worse than that of WTR astigmatism, and the image reliability of eyes with ATR astigmatism was better than that of WTR astigmatism.

There were several limitations in our study. First, due to the screening process, we were unable to collect comprehensive sociodemographic information from the participants. Second, we did not grade the severity of media opacity, particularly cataracts, which could have affected image quality. Additionally, we did not perform dilated fundus photography, as this was not feasible in a large-scale community-based screening with a high population density. Furthermore, it should be noted that subjects with narrow angle anatomy may be at risk of acute angle-closure glaucoma attacks when their pupils are dilated. Third, to provide timely feedback regarding diagnosis during onsite screening, we did not conduct hierarchical image reading. However, the reading staff had more than 10 years of experience in retina reading, and the interobserver and intraobserver agreement was moderate to high. Finally, due to a large number of individuals usually making appointments for health examinations everyday during the screening period, we did not investigate different photographic conditions on image quality, such as photography intervals, exposure intensity and so on. Meanwhile, recently different fundus cameras such as ultra-wide fundus imaging, retro-mode imaging based on confocal scanning laser ophthalmoscopy and fundus adaptive optics scanning laser ophthalmoscope emerged rapidly. These cameras greatly improved the quality and resolution of fundus imaging compared with traditional fundus cameras, especially for old population with media opacity, and without pharmacological pupil dilation. These fundus imaging techniques might improve the image reliability in the real world.

Our study suggests that single-field NMFPs can be used for large-scale glaucoma screening in the general population over 50 years old, but the reliability of the images may be compromised. Age, absolute value of SE, media opacity and VA are significant factors that affect image quality. Oblique astigmatism is associated with worse image readability than WTR astigmatism, while ATR astigmatism is associated with better image reliability than WTR astigmatism. These findings highlight the importance of addressing these factors to improve the effectiveness of ophthalmological telemedicine and to optimise the screening mode for eye diseases in underserved areas.