Monitoring glaucoma progression from Fundus photos using a deep learning‐based trend analysis

Aims/Purpose: To assess the agreement between established tools to monitor glaucoma progression and a linear regression on deep learning (DL)‐based glaucoma risk predictions obtained from color fundus photos (CFPs).

Methods: CFPs and clinical metadata were retrospectively collected from three university hospitals in Belgium. Eyes with at least five CFP visits were included based on the availability of progression ground truth labels.

Ground truth labels for glaucoma progression were derived from within‐eye linear regression on either average retinal nerve fiber layer thickness (aRNFL) as measured by optical coherence tomography (OCT) or mean deviation (MD) from visual field testing. A significant MD slope (p < 0.05) from linear regression was selected as a biomarker for functional glaucoma progression, tested with a minimum of 4, 5, and 6 values. For structural progression, a significant aRNFL slope (p < 0.05) was assessed, requiring 3 or 4 visits due to data scarcity and lower test‐retest variability in OCT data.

Preprocessed CFPs were input into a previously trained deep learning model (G‐RISK). The within‐eye linear regression of G‐RISK predictions, referred to as the G‐RISK slope, was then obtained. Agreement was evaluated using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals. We varied the MD slope between 0 and ‐4 dB per year and computed AUC using G‐RISK slopes as predictions. Similarly, we varied the aRNFL slope between 0 and ‐4 μm per year.

Results: A total of 5875 eyes had at least four visits with MD values, and 313 eyes had at least three visits with aRNFL values. For functional progression, AUC climbed up to 0.67 [0.58; 0.75] at an MD slope of at least ‐2.5 dB per year. For structural progression, AUC reached up to 0.99 [0.98; 1] with a yearly aRNFL decrease exceeding 3.4 μm.

Conclusions: Both MD and RNFL slopes showed significant agreement with G‐RISK slope, supporting its potential as a novel biomarker for glaucoma management.

References