Glaucoma Progression Detected Using a Deep Learning Model

A deep learning model shows good agreement with clinical experts and outperforms a trend-based analysis for identifying glaucoma progression.

A deep learning model may predict the probability of glaucomatous damage using retinal nerve fiber layer (RNFL) thickness obtained from spectral domain-optical coherence tomography (SD-OCT), according to research published in Ophthalmology Glaucoma. The model demonstrated good agreement with glaucoma experts and outperformed a trend-based analysis, according to the report. 

Researchers performed a retroscopic record review of 14,034 SD-OCT scans of 816 eyes (glaucoma, 446; glaucoma suspects, 129; control eyes, 241) from 462 individuals (mean age, 64.5 years; 58.4% women). All participants had at least 1 year of follow-up and 3 SD-OCT scans. Two glaucoma specialists evaluated the SD-OCT images, determined whether glaucoma progression had taken place, and identified the visit of initial progression in patients who experienced progression. 

Patients’ RNFL thickness profiles were input into the deep learning model along with the time period between initial and follow-up visits. The team evaluated the deep learning model’s accuracy by its ability to differentiate between scans that were stable vs those that indicated progression according to the glaucoma specialists. Diagnostic performance was evaluated using area under the receiver operator characteristic curve (AUC), sensitivity, and specificity analysis. 

The AUC was 0.938 (95% CI, 0.921-0.955), demonstrating an ability to differentiate between stable and progressing glaucoma. The sensitivity was 87.3% (95% CI, 83.6%-91.6%) and specificity was 86.4% (95% CI, 79.9%-89.6%). The deep learning model outperformed a trend-based analysis, which demonstrated 46.1% sensitivity (95% CI, 36.7%-55.0%) and 92.6% specificity (95% CI, 90.7%-94.3%) using global RNFL thickness. A trend-based analysis that considered both global and sectoral RNFL thicknesses demonstrated a sensitivity of 83.7% (95% CI, 78.4%-88.7%) and specificity of 68.6% (95% CI, 65.5%-71.7%). The deep learning model’s sensitivity was 96.2% (95% CI, 94.3%-97.8%) at this level of specificity with a significant absolute difference of 12.5% (95% CI, 7.6%-17.8%).

Probabilities of progression provided by the model along with the visualization heatmaps may help clinicians in identifying structural glaucoma progression with SD-OCT. 

“The [deep learning] model agreed well with expert judgments of progression and outperformed conventional trend-based analyses,” according to the researchers. “Probabilities of progression provided by the model along with the visualization heatmaps may help clinicians in identifying structural glaucoma progression with SD-OCT.”

Study limitations include a lack of visual field testing to assess visual functioning.

Disclosure: Several study authors declared affiliations with biotech, pharmaceutical, and/or clinical research organizations. Please see the original reference for a full list of authors’ disclosures.


Mariottoni EB, Datta S, Shigueoka LS, et al. Deep learning assisted detection of glaucoma progression in spectral-domain optical coherence tomography. Ophthalmol Glaucoma. Published online November 17, 2022. doi:10.1016/j.ogla.2022.11.004