Abstract
Purpose :
In diabetic macular edema (DME), hard exudates (HE) represent exudative material and other deposits in the retina, and are one of the hallmarks of diabetic retinopathy (DR). We applied a deep learning-based (DL) segmentation model for HE in eyes with DME using spectral domain optical coherence tomography (OCT) scans to measure HE on a volumetric level, describe model performance with results on baseline data, and association to human color fundus photography (CFP)-based grading.
Methods :
Hyperreflective material was annotated on the pixel level on B-Scans from the Phase 2 BOULEVARD study (NCT02699450) by 2 experts from Liverpool Ophthalmic Reading Centre, split on the patient level into training (1355 B-Scans) and validation (155 B-scans) sets. The U-Net, a convolutional neural network for biomedical image segmentation, was trained and evaluated using Sørensen–Dice coefficient (DICE) scores, measuring the overlap between annotations and predictions, and specificity. 98 OCT Spectralis scans with 97 B-Scans each from the screening visit of the Phase 2b ALTIMETER study (NCT04597918) were segmented and objects > 100 µm in diameter were classified as HE. HE counts and volumes were assessed, and volumes compared to manual grading on CFP, where “definite” or “absent” encoded HE presence/absence (n=3 “questionable” cases excluded).
Results :
On the validation set the model reached median and average DICE scores of 71% and 65%, respectively, with a median specificity of 99.96%, ruling out noise detection. In the 3 mm diameter ETDRS ring, baseline HE median volume was 4.7 nl, with the first quartile at 0.71 nL and the third quartile at 41.18 nL (IQR=40.47). Baseline HE median object count was 5.5, with the first quartile at 2 and the third quartile at 18 (IQR=16). The segmented median HE volume was 70 times higher (26.8 nl vs 0.38 nl) in the group of eyes graded “definite” (n=68) vs “absent” (n=27). A high degree of spatial concordance between HE on CFP vs HE on OCT en-face projections was observed, which will be shown in the presentation.
Conclusions :
We developed a DL-based segmentation model for HE on OCT, and assessed its performance. We applied it to baseline scans of DME eyes, assessed HE volume and count distributions, and showed good alignment with CFP-based binary presence/absence by the reading centre, enabling quantitative volumetric assessment of HE on OCT.
This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.