Foundation Models Permit Retinal Layer Segmentation Across OCT Devices

The segmentation of retinal layers in images from optical coherence tomography (OCT) is an important step in ophthalmological diagnosis and disease monitoring. Current CNN-based models perform well on images from the same OCT scanner on which they have been trained, but their performance can degrade drastically when images are acquired with other devices. We present the first method for OCT layer segmentation that builds on recent Vision Transformer (ViT) foundation models. We demonstrate that, compared to a state-of-the-art CNN approach, doing so significantly improves their ability to generalize to devices for which no training data was available. This highlights the potential of foundation models to enable more robust medical image analysis. We also analyze the effect of using different foundation models. Notably, more generic foundation models from computer vision permitted better generalization than an equally large foundation model that was specifically trained for OCT analysis.