Deep learning-based optical coherence tomography segmentation of retinal fluids and layers using multi-grader annotations

Abstract

Purpose : To use multi-grader annotations for developing artificial intelligence (AI)-based retinal fluid and layer
segmentation models for optical coherence tomography (OCT) scans, to enable disease activity
quantification and monitoring in neovascular age-related macular degeneration (nAMD) and diabetic
macular edema (DME).

Methods : The dataset used to develop and validate the segmentation models contained 2880/5227 B-scans from
493/515 SD-OCT volumes of 229/276 patients from clinical trials (NCT02484690, NCT02699450,
NCT02510794, NCT03038880, NCT04597918) for retinal fluids/layers, respectively. Most volumes were
from screening and first visits (51.5%) to ensure images with maximum pathological fluid biomarkers were
included; others came from follow-up visits after treatment. A sparse selection of 5-33 B-scans was
annotated per volume for 6 biomarkers (intraretinal fluid [IRF]; cystoid spaces), subretinal fluid [SRF],
pigment epithelial detachment [PED], internal limiting membrane [ILM], inner boundary-retinal pigment
epithelium [IB-RPE], and Bruch’s membrane [BM]), clearly defined to ensure annotation
quality/consistency. Each B-scan was annotated by a random selection of 4 of 8 participating graders.
The data were split into training/validation (90%/10%) sets at the patient level.
Multi-grader annotations were aggregated into a consensus annotation using the STAPLE algorithm
before being fed into a convolutional neural network, a 2D U-Net, for training the segmentation models.
Models were evaluated using the validation set and compared with aggregated multi-grader annotations
using the Sørensen–Dice coefficient and Chamfer Distance for fluids and layers, respectively.
Generalized Bland-Altman plots analyzed the similarity between the model and human performance.

Results : Median Dice scores were 0.816/0.898/0.887 for IRF/SRF/PED. Median Chamfer Distances (μm) were
7.689±2.543/2.125±2.677/6.780±4.180 for ILM/IB-RPE/BM. Generalized Bland-Altman plots (Fig. 1)
showed that the algorithm strongly agreed with the human graders.

Conclusions : Algorithm validation analyses showed that the performance of the AI-based OCT segmentation model
was similar to that of expert graders, which enables fast, automated and reliable OCT image analysis that
is important for monitoring/managing nAMD and DME.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.