Predicting optical coherence tomography-derived diabetic macular edema grades from fundus photographs using deep learning

Introduction

• OCT is used to assess macular thickening for DME treatment eligibility, but adding OCT to screening is costly and logistically difficult to implement widely.Most clinical guidelines recommend annual screening for people with diabetes.Source: Introduction · p. 3
• Hard exudates within one optic disc diameter of the macular center serve as a proxy for ci-DME, but studies report poor positive predictive value and sensitivity; only 42% of patients with hard exudates had DME on OCT.Source: Introduction · pp. 3–4
• Deep learning has been applied to retinal-image tasks, and the authors hypothesized it could predict OCT-derived DME grades using monoscopic fundus photographs.The study frames this as a way to address unreliable ci-DME detection from fundus photographs alone.Source: Introduction · p. 4

Results

• 85% sensitivity at 80% specificity for ci-DME exceeded specialists’ 82-85% sensitivity at 45-50% specificity; at matched specificity, sensitivity was 96% versus 82-85%.At the same sensitivity, the model also had 30-35% higher absolute specificity than manual grading (p<0.001 for each specialist comparison).Source: Results · p. 5
• AUC was 0.81 (95% CI: 0.81-0.86) for intraretinal fluid and 0.88 (95% CI: 0.85-0.91) for subretinal fluid.Source: Results · pp. 5–6
• Absolute performance across the primary and secondary validation datasets is not directly comparable, especially for PPV, because ci-DME definitions and prevalence differed.The secondary dataset had 7.8% ci-DME cases versus 27.2% in the primary validation set.Source: Results · p. 6
• Model AUC increased with larger training sets, suggesting accuracy may continue to improve with datasets larger than this study’s.Source: Results · p. 7
• A 0.25 optic-disc-diameter crop around the fovea yielded AUC 0.75, while a 1.0-diameter foveal crop exceeded 0.85 and was comparable to full-fundus performance; optic-disc crops required 2.5 diameters of context.Source: Results · p. 7

Discussion

• 88% of EyePACS-DME and 77% of Thailand moderate non-proliferative DR patients urgently referred by hard-exudate criteria lacked ci-DME.These would have been unnecessary urgent referrals under the hard-exudate-based criterion.Source: Discussion · p. 9
• The model’s false positives and false negatives were more often borderline cases than retina specialists’ errors, including 17% of false negatives with thickness below 275 microns.The passage also reports false positives with thickness above 225 microns, but does not provide their proportion.Source: Discussion · p. 9
• The model appears able to detect intraretinal and subretinal fluid, pathologies that merit closer monitoring and possibly treatment and that doctors cannot accurately identify from fundus images.The authors describe this fluid-detection capability as novel.Source: Discussion · p. 9
• 35% was the model’s secondary-dataset PPV, down from 61% on the primary dataset; graders’ PPV fell from 37% to 18%, while sensitivity fell from 85% to 57% for the model and 84% to 55% for graders.The authors report that model performance was lower on the secondary dataset, but graders’ performance declined proportionally as well.Source: Discussion · p. 9
• 96% was the model’s secondary-dataset NPV, up from 93% on the primary dataset; graders’ NPV rose from 88% to 95%, while specificity rose from 80% to 91% for the model and 47% to 79% for graders.The authors say these results reflect inherent dataset differences but still support better model performance than graders on both datasets.Source: Discussion · p. 10
• Cropping experiments and the explanation technique indicated that the region around the fovea is more relevant than the optic-disc region for predicting DME from fundus images.The discussion presents these analyses as groundwork for understanding how the model makes predictions.Source: Discussion · pp. 10–11
• The study demonstrates that deep learning can identify ci-DME from inexpensive, widely available fundus photographs with accuracy exceeding manual grading using expert-derived rules.Secondary-dataset performance was lower, confidence intervals were wide, and ci-DME criteria differed between datasets; PPV and NPV apply to populations with a similar severity distribution, and further validation across settings is needed. The model does not evaluate cases without central involvement that meet older treatment criteria.Source: Discussion · pp. 10–11

Measurement and assessment of OCT scans

• Intraretinal fluid was defined by a hypo-reflective cystoid space within 500 μm of the foveal center, and subretinal fluid by a hypo-reflective space between retina and retinal pigment epithelium there.Two experienced medical doctors assessed OCT measurements and fluid presence under retinal-specialist supervision; a specialist reviewed a random 5% of patients.Source: Measurement and assessment of OCT scans · p. 14
• ci-DME was defined as center point thickness >= 250 μm excluding the retinal pigment epithelium; EyePACS-DME used a machine-adjusted 300 μm central subfield thickness cutoff.Intraretinal and subretinal fluid labels were unavailable for EyePACS-DME.Source: Measurement and assessment of OCT scans · pp. 14–15
• The model was trained to predict ci-DME, intraretinal fluid, and subretinal fluid from fundus photographs, without seeing OCT images during training or validation.The model output a confidence value for each prediction.Source: Measurement and assessment of OCT scans · p. 15

Evaluating the algorithm

• The model was evaluated with ROC curves and AUC, specialist operating points, and additional thickness thresholds; metrics were also assessed at an operating point matching specialists’ specificity.The reported metrics included sensitivity, specificity, PPV, NPV, accuracy, and Cohen’s kappa.Source: Evaluating the algorithm · p. 16

Statistical Analysis

• Statistical significance was assessed with 2,000 bootstrap resamples for 95% confidence intervals and 2,000 random permutations for comparisons.Bootstrap confidence intervals used the 2.5th and 97.5th percentiles of the resampled metric distribution.Source: Statistical Analysis · p. 16

Model Explanation

• To test which image regions inform DME predictions, models were evaluated on progressively larger fovea-centered regions, with the rest of each image blacked out.The fovea and optic disc were manually marked, and region size was varied by optic-disc diameters.Source: Model Explanation · pp. 16–17

Data availability

• Study data may be shared with qualified researchers following ethical approvals from Rajavithi Hospital.The authors state that restrictions apply to sharing the patient data supporting the findings.Source: Data availability · p. 17

Code availability

• Models were developed and deployed with standard libraries and TensorFlow; custom infrastructure-specific code mainly handled data input/output and parallelization across computers.Source: Code availability · p. 18

• The authors are employees of Google.Source: Competing interests · p. 18
• The model and retinal specialists were compared on the primary validation set using OCT-defined ci-DME at a 250 μm center-point threshold.Source: Figure 2 · p. 21
• 85% sensitivity at 80% specificity was the model’s ci-DME operating point on the primary clinical validation set.Source: Table 2 · p. 22
• Model performance increased with training-data size across predictions of ci-DME, subretinal fluid, and intraretinal fluid.Source: Figure 3 · p. 23
• The secondary validation compared the model with EyePACS graders using an OCT central-subfield threshold of 300 μm.Source: Figure 4 · p. 24
• The crop analysis compares AUC across circular image regions centered on the fovea or optic disc, with radii from 0.05 to 2.5 disc diameters.Source: Figure 5 · p. 25

C. CPT>=320 μm

• OCT center-point thickness cutoffs of 280, 300, and 320 μm were used in supplementary ci-DME evaluations.Source: Figure S2 · p. 30
• False-positive and false-negative center-point thickness distributions were compared for the model and specialists at an operating point matching specialist sensitivity.Source: Figure S3 · p. 31
• The supplementary materials also identify the training data augmentation and model hyperparameters and compare the Thailand and EyePACS-DME datasets.Source: Figure S4 · p. 32, Table S2 · p. 33
• Table S1 reports performance metrics on 948 images all three retinal specialists deemed gradable, with the model operating point matched to specialist sensitivity.Source: Table S1 · p. 33