Abstract
Purpose :
In diabetic macular edema (DME), hyperreflective foci (HRF) may represent an inflammatory response within the retina. The purpose of this study was to validate a deep learning-based (DL) segmentation model for HRF in spectral domain optical coherence tomography (SD-OCT) volumes in DME to measure HRF volumes automatically.
Methods :
334 Spectralis (Heidelberg Engineering) SD-OCT volumes from the phase 2 BOULEVARD study (NCT02699450) were selected. Per volume, 2-9 B-scans were annotated on the B-scan level by 2 experts from Liverpool Ophthalmic Reading Centre, by placing ellipses over HRF up to 50 µm in diameter, and manually outlining larger hyperreflective objects. Objects with lower reflectivity than Retinal Pigment Epithelium (RPE) were not annotated. Ellipses were shrunk to the most hyperreflective center using adaptive thresholding. Data was split on the patient level into training (1355) and validation (155) image sets. The U-Net, a convolutional neural network for biomedical image segmentation, was trained for 250 epochs. Model performance was evaluated on the validation set using Sørensen–Dice coefficient (DICE) scores, measuring the overlap between annotations and model predictions.
The trained model was applied to all BOULEVARD volumes, and detected objects were categorized by fitting an ellipsoid. Objects below 50 µm in diameter (long axis) were classified as HRF and volumes automatically extracted.
Results :
Median and average DICE scores on the validation set were 71% and 65%, respectively. Baseline absolute HRF volumes (n=217) were assessed in the central 1.0 and 3.0 mm diameter cylinder slabs between Internal Limiting Membrane and Outer Plexiform Layer/Henle’s Fiber Layer (OPL/HFL), with Median (Interquartile Range) values of 120,932 (341,832) µm3 and 1,511,948 (1,869,309) µm3, respectively. The corresponding values between OPL/HFL and RPE were 67,304 (262,643) µm3 and 982,169 (2,099,866) µm3, respectively.
Conclusions :
Exhaustive manual annotation of HRF is unfeasible. A DL-based segmentation model using a sparse subset of annotated B-Scans enables fully automatic and reliable HRF segmentation in eyes with DME. In the future, this model will enable large-scale evaluation of changes in HRF with treatment, and help elucidate the clinical meaning of HRF.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.