GANSeg

GANSeg performs unsupervised cross-domain segmentation of macular optical coherence tomography (OCT) B-scans to identify seven retinal layers and intraretinal fluid, enabling models trained on one OCT device to generalize to other manufacturers without labeled target-device data.

Key Features:

• Unsupervised cross-domain training: Trained using labeled Heidelberg Spectralis images (domain A) and applied to unlabeled Topcon 1000 images (domain B) within a dataset of 732 OCT B-scans from Heidelberg Spectralis, Topcon 1000, Topcon Maestro2, and Zeiss Plex Elite 9000.
• Generative adversarial network (GAN): Uses a GAN-based approach for domain adaptation to bridge distributional differences between OCT devices without labels on the target device.
• Segmentation targets: Produces segmentation of seven retinal layers and intraretinal fluid in Topcon 1000 OCT B-scans.
• Validation against human graders: Performance was validated on an external Topcon 1000 test dataset using manual segmentations from three masked graders.
• Reported performance metrics: Achieved Dice scores comparable to human experts for certain layers, including 90% (95% CI, 68%–96%) for the ganglion cell layer plus inner plexiform layer and 58% (95% CI, 18%–89%) for intraretinal fluid segmentation with statistical comparisons to graders reported.
• Generalization and baseline comparison: Extended to Zeiss Plex Elite 9000 and Topcon Maestro2 without prior exposure and outperformed a baseline U-Net trained on the same Heidelberg images.

Scientific Applications:

• Cross-device retinal image analysis: Enables segmentation-based studies that integrate macular OCT data from multiple manufacturers without requiring labeled datasets for each device.
• Quantitative retinal layer and fluid assessment: Facilitates measurement of retinal layer morphology and intraretinal fluid for ophthalmic research and comparative studies.
• Evaluation of automated segmentation versus human graders: Supports validation studies comparing automated segmentation performance to masked expert manual annotations using Dice-based metrics and confidence intervals.

Methodology:

GANSeg employs a generative adversarial network for unsupervised domain adaptation, trained on labeled Heidelberg Spectralis images and applied to unlabeled Topcon 1000 images using a dataset of 732 OCT B-scans from four devices, with validation against manual segmentations by three masked graders on an external Topcon 1000 test set.