Abstract
Purpose :
Deep learning–based segmentation models for spectral-domain optical coherence tomography (SD-OCT) images have been used for automatic identification of retinal fluids to support clinicians in patient management. Images from different acquisition devices or retinal diseases, hereafter referred to as domains, typically require distinct models. We propose an effective unified solution applicable to multiple domains.
Methods :
SD-OCT images were acquired with Cirrus HD-OCT (Carl Zeiss Meditec), denoted as C, and Spectralis OCT (Heidelberg Engineering), denoted as S, on patients with neovascular age-related macular degeneration (nAMD) and diabetic macular edema (DME). The dataset was grouped into 3 domains: S/nAMD (AVENUE, NCT02484690), S/DME (BOULEVARD, NCT02699450), and C/nAMD (HARBOR, NCT00891735), which respectively included 1769, 1686, and 959 B-scans (corresponding to 303, 303, and 160 eyes) with manual pixel-wise annotations. The annotations denote IRF and SRF for patients with DME, as well as PED and SHRM for patients with nAMD (Figure 1). Annotated B-scans from 10% of eyes in each domain were randomly split into a test set for evaluation.
Different domain-specific models with a UNet architecture were trained on each of the 3 domains as a baseline (Figure 1, left). Our cross-domain model was trained collectively on B-scans from all 3 domains (Figure 1, right). Experiments were repeated 5 times to minimize the effect of randomness in network initialization. Performance was measured using the Dice coefficient.
Results :
Domain-specific models were confirmed to be superior (5.68 higher Dice overall) when training and evaluation domains were the same (Figure 2). Segmentation results from our cross-domain model were comparable (3/10 cases) or superior (7/10 cases) to any domain-specific alternative for all domains and fluid labels (2.43 higher Dice overall). Training of the cross-domain model was also found to be > 2.5× faster.
Conclusions :
Overall, our proposed cross-domain segmentation model is more accurate and faster to train compared with domain-specific models and eliminates the need for customizing models domain-specifically. The benefits of our approach may facilitate automated SD-OCT segmentation in the clinics.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.