Monitoring glaucoma progression from fundus photos using a novel deep learning-based trend analysis

Abstract

Purpose : To assess the agreement between a human-generated glaucoma progression ground truth (GT) and a linear regression on deep learning (DL)-based glaucoma risk predictions obtained on individual photos.

Methods : Data were retrospectively collected from the population-based Blue Mountains Eye Study (BMES, Australia) and from the clinical records of Tays Eye Centre (TEC, Finland). Inclusion criteria were defined as having at least three visits with fundus photos of sufficient quality, collected over a period between 2 and 10 years. The image quality was objectively scored by an auxiliary DL model.

For BMES, the human progression GT reflects the conversion from non-glaucoma to glaucoma between the baseline (year 0) and either follow-up visit (year 5 or 10). See Fig 1 for an eye with glaucoma onset.

For TEC, the human progression GT is available for patients that visited the clinic in 2019 and/or 2020. The human expert assessed the presence of glaucoma-induced changes to optic nerve head (ONH), retinal nerve fiber layer (RNFL) and visual field (VF) from either retinal images or perimetry.

Fundus photos served as the input for a ResNet-50 model previously trained for the estimation of vertical cup-disc ratio values, denoted as G-RISK. We performed Ordinary Least Squares (OLS) regression on G-RISK predictions for a specific eye, with the visit year as the explanatory variable. The resulting slope from the OLS output (G-RISK slope) was then compared to the human-generated GT by using the area under the receiver operating characteristic curve (AUC). A sensitivity analysis was conducted using the maximum G-RISK baseline value.

Results : The eligible sets consisted of 964 and 1570 eyes in BMES and TEC, respectively. The median time between visits was fixed in BMES (5 years), and irregular in TEC (1.75 years).

In BMES, G-RISK slope agreement with human GT varied from 0.54 to 0.88 (AUC), depending on the upper limit on baseline G-RISK. Similar behavior was observed in the TEC data, with AUC ranging between 0.54 and 0.76. The highest AUC values were recorded in the ONH GT, followed by VF and RNFL (Fig 2).

Conclusions : G-RISK slope obtains good agreement on conversion to glaucoma in two distinct data sets with two independent human ground truths. Trend analysis on deep learning-based predictions could be useful in the monitoring of a non-glaucomatous population to detect conversion to glaucoma.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Fig 1

 

Fig 2