Deep Learning Model Automates AMD Segmentation Features

Researchers developed a deep learning model for segmentation of 13 features commonly found in neovascular and atrophic age-related macular degeneration (AMD) that, they said, performed at a level comparable, and “for some features possibly better,” than independent observers. They published their work in American Journal of Ophthalmology.

The researchers sought to validate the quality of their model’s automatic segmentation of features commonly found in neovascular and atrophic AMD in an effort to provide an alternative to manual volumetric quantification of parameters seen on optical coherence tomography (OCT).

The researchers selected 13 common annotated abnormalities, or features, to be included in the development of the model: intraretinal fluid (IRF), subretinal fluid (SRF), pigment epithelial detachment (PED), subretinal hyperreflective material (SHRM), fibrosis, drusen and drusenoid PED (Drusen), epiretinal membrane (ERM), outer plexiform layer descent (OPL descent), ellipsoid loss, retinal pigment epithelium loss or attenuation (RPE loss), hypertransmission (HTR), hyperreflective dots and exudates (HRD) and subretinal drusenoid deposits-reticular pseudodrusen (SDD-RPD).

They scanned the eyes of 30,337 patients presenting with neovascular AMD at 5 United Kingdom centers using OCT and created 1 set of 2712 annotated B-scans from 307 OCT scans of 307 eyes for model development, which 8 graders manually delineated, and another 1 in which 4 graders delineated all abnormalities in a single B-scan per OCT volume, to assess agreement between graders and model. The test set included 112 OCT scans from 112 eyes, independent of the development data. Half were obtained before the patient’s first anti-vascular endothelial growth factor (VEGF) injection, and the others were acquired about 3 months after the first injection.

The model obtained a Dice-score of 0.63 (± 0.15, median 0.64), compared with 0.61 (± 0.17, median 0.60) for the observers. The average intra-class correlation coefficient (ICC) score for the model was 0.66 (± 0.22, median 0.69), compared with 0.62 (± 0.21, median 0.55) for the observers. Differences between model and observer Dice-score were within a 95% confidence interval for all features except ellipsoid loss, where model performance was higher (P =.03). The model performance was higher for IRF (P =.04) and ellipsoid loss (P =.006), lower for drusen (P =.03), and within the 95% confidence interval of the observer score for the other features. For HRD, which was assessed using free-response receiver operating characteristic (FROC) curve, the model obtained a sensitivity higher than each grader when operating at the same false positive rate. The model detected complete RPE and outer retinal atrophy (cRORA) in 13 B-scans (6 of the 7 B-scans with cRORA and 3 of the 96 B-scans without cRORA), and 3 of the 4 graders reached consensus that cRORA was present in 7 of the 112 B-scans, absent in 96 B-scans and ambiguous in 9 B-scans.

Want to read more?