Source-linked AI summary
From Machine to Machine: An OCT-trained Deep Learning Algorithm for Objective Quantification of Glaucomatous Damage in Fundus Photographs
Felipe A. Medeiros, Alessandro A. Jammal, Atalie C. Thompson
TL;DR
Subjective grading of optic disc photographs is poorly reproducible and can misestimate glaucoma likelihood, while human-labeled training data may reproduce those errors. This study trained deep learning on SDOCT RNFL measurements to quantify glaucomatous damage from photographs, achieving strong agreement with SDOCT measurements and similar glaucoma discrimination.
Problem
Subjective optic disc-photo evaluation has low reproducibility, and human-labeled training may replicate grading errors, motivating objective training targets.
Method
A deep learning algorithm was trained on optic disc photographs paired with SDOCT average RNFL thickness to produce continuous estimates of neural damage.
Results
r = 0.832 correlation between predicted and observed RNFL thickness, with 7.39 μm mean absolute error; ROC areas were 0.944 for predictions and 0.940 for actual SDOCT measurements.
Takeaways & Limitations
The approach provides quantitative information from optic disc photographs and could potentially support diagnosing, staging, and monitoring glaucomatous damage where SDOCT is unavailable.
Takeaways & Limitations
Approximately 30% of SDOCT measurement variance remained unexplained, and photograph quality was not qualitatively assessed.
Abstract
from arXiv · showhide
Previous approaches using deep learning algorithms to classify glaucomatous damage on fundus photographs have been limited by the requirement for human labeling of a reference training set. We propose a new approach using spectral-domain optical coherence tomography (SDOCT) data to train a deep learning algorithm to quantify glaucomatous structural damage on optic disc photographs. The dataset included 32,820 pairs of optic disc photos and SDOCT retinal nerve fiber layer (RNFL) scans from 2,312 eyes of 1,198 subjects. A deep learning convolutional neural network was trained to assess optic disc photographs and predict SDOCT average RNFL thickness. The performance of the algorithm was evaluated in an independent test sample. The mean prediction of average RNFL thickness from all 6,292 optic disc photos in the test set was 83.3$\pm$14.5 $μ$m, whereas the mean average RNFL thickness from all corresponding SDOCT scans was 82.5$\pm$16.8 $μ$m (P = 0.164). There was a very strong correlation between predicted and observed RNFL thickness values (r = 0.832; P<0.001), with mean absolute error of the predictions of 7.39 $μ$m. The areas under the receiver operating characteristic curves for discriminating glaucoma from healthy eyes with the deep learning predictions and actual SDOCT measurements were 0.944 (95$\%$ CI: 0.912- 0.966) and 0.940 (95$\%$ CI: 0.902 - 0.966), respectively (P = 0.724). In conclusion, we introduced a novel deep learning approach to assess optic disc photographs and provide quantitative information about the amount of neural damage. This approach could potentially be used to diagnose and stage glaucomatous damage from optic disc photographs.
1 Introduction
Glaucoma requires objective structural assessment because fundus photographs are accessible but subjective grading is poorly reproducible. This study trains deep learning on SDOCT measurements to quantify glaucomatous neural loss from optic disc photographs.
- Early structural assessment matters because treatment can prevent or slow glaucoma’s progression to functional damage and visual impairment.
- Fundus photographs are low-cost and easy to perform, but expert subjective grading has low reproducibility and can misestimate glaucoma likelihood.
- Prior deep learning classified referable glaucomatous optic neuropathy using photograph labels based on subjective ophthalmologist grading.
- Because algorithms cannot exceed their training reference standard, human-labeled training may reproduce subjective grading errors.
- SDOCT RNFL thickness provides a reproducible, objective reference standard that accurately detects glaucoma, progression, and rates of change.
- The study introduces a deep learning algorithm trained from SDOCT results to assess optic disc photographs and quantify glaucomatous neural loss.
2 Methods
The study paired longitudinal optic disc photographs with nearby SDOCT RNFL scans, then trained and evaluated a patient-level separated deep learning model. Analyses compared predicted with observed RNFL thickness and assessed diagnostic discrimination.
- Data source: The repository combined ophthalmologic records, stereoscopic optic disc photographs, visual fields, and Spectralis SDOCT images and data.
- Data acquisition: SDOCT RNFL scans acquired 1536 A-scan points around a 3.45-mm circle centered on the optic disc.
- Pairing: Each eye’s available optic disc photographs were matched with the closest SDOCT RNFL scan acquired within 6 months of the photograph.
- Pairing: Multiple longitudinal photograph–SDOCT pairs increased dataset heterogeneity for deep learning training.
- Model development: The network predicted SDOCT average RNFL thickness from optic disc photographs using an 80% training-validation split and a 20% test sample, sampled at the patient level to prevent leakage.
- Preprocessing: Photographs were split into stereo views, downsampled to 256 x 256 pixels, scaled from 0 to 1, and augmented to reduce overfitting.
- Model development: The model used a pretrained ResNet34, progressively unfroze layers, and trained with differential learning rates, minibatch gradient descent, and Adam optimization.
- Statistical analysis: Analyses evaluated continuous RNFL predictions with mean absolute error, Pearson correlation, Bland–Altman agreement, and ROC discrimination, accounting for repeated patient measurements.
3 Results
In an independent test sample, the algorithm’s optic-disc-photo predictions closely tracked SDOCT RNFL measurements and differentiated normal, suspect, and glaucomatous eyes. Its diagnostic performance was comparable to actual SDOCT measurements, while activation maps emphasized the optic disc and adjacent RNFL.
- 6,292 test pairs yielded mean predicted RNFL thickness of 83.3±14.5 µm versus 82.5±16.8 µm for corresponding SDOCT scans (P = 0.164).The prediction error was 7.39 µm on average, with Pearson’s r = 0.832 and 95% agreement limits from -18.5 to 17.5 µm.
- Predicted RNFL thickness differed significantly across normal, suspect, and glaucomatous eyes, with means of 96.1±7.8, 87.5±9.9, and 71.0±14.4 µm, respectively (P<0.001).Corresponding SDOCT means were 97.6±9.3, 87.1±12.5, and 68.8±16.0 µm, with P<0.001 for all pairwise comparisons.
- Predicted RNFL thickness correlated with visual field mean deviation at r = 0.61, similar to actual SDOCT RNFL thickness at r = 0.59.The corresponding R2 values were 37% for predictions and 35% for actual SDOCT measurements.
- The ROC area was 0.944 for deep-learning predictions versus 0.940 for actual SDOCT RNFL thickness, with no significant difference (P = 0.724).At 80% specificity, both approaches achieved 90% sensitivity.
- The network achieved 83.7% accuracy against SDOCT normative-database classifications.Class activation maps concentrated most strongly on the optic nerve and adjacent RNFL, while retinal blood vessels and more distant regions had smaller activations.
4 Discussion
The study replaces subjective human labels with SDOCT-derived RNFL measurements to train a deep learning model that quantitatively assesses glaucomatous damage from optic disc photographs. Predictions closely tracked measured RNFL thickness and showed similar discrimination of glaucoma from healthy eyes, while limitations included unexplained variance and the need for longitudinal validation.
- Novel quantitative approach: Continuous RNFL thickness predictions were generated from optic disc photographs by training the network with SDOCT measurements rather than human labels.This approach was presented as a novel way to quantify neural damage on disc photographs.
- Rationale: Human-labeled training sets can reproduce subjective grading errors, including misclassifying physiologic large cups and missing glaucoma in small optic discs.Objective and reproducible RNFL thickness measurements provide an alternative target and avoid time-consuming subjective labeling.
- Performance: The model’s predictions showed very strong correlation with actual RNFL thickness in the independent test sample, with a mean absolute error of approximately 7 µm.The correspondence supported validation against clinically relevant structural measurements.
- Performance: Predicted RNFL thickness and actual SDOCT measurements produced nearly identical ROC curve areas for discriminating glaucomatous visual field loss from healthy eyes.The discussion describes this similarity as important validation of the model.
- Potential utility: Fundus photographs may offer a quicker, less expensive, and more portable alternative to SDOCT for documenting optic disc appearance, especially in screening settings.SDOCT machines are described as expensive and non-portable, whereas photographs may be easier to implement.
- Potential utility: Continuous outputs could potentially support monitoring of glaucomatous change over time where SDOCT is unavailable, but this use requires longitudinal validation.A binary assessment of photographs does not generally provide progression information.
- Model interpretation: Activation heatmaps localized the model’s most important image regions to the optic disc and adjacent RNFL rather than retinal blood vessels or more distant areas.The authors suggest this may help identify difficult disc features and improve clinician training.
- Limitations: Approximately 30% of SDOCT measurement variance remained unexplained, with possible contributions from disc appearance, SDOCT variability, and photograph quality.The algorithm predicted average rather than segmental RNFL thickness, and larger datasets may improve performance.