Zero-shot OCT segmentation for volumetric and live-streaming data

Deep-learning-based OCT segmentation methods are accurate but less flexible (need re-labeling and re-training) for unseen and varying Regions Of Interest (ROIs), which hinders their applications in many scenarios, such as surgical and navigation. Inspired by the recent emerging large vision models and their zero-shot capabilities (directly applicable to unseen ROIs), we propose a zero-shot OCT segmentation method. For volumetric and live-streaming data, it only requires the box or point prompts of ROIs at the first frame, then subsequent frames could be segmented automatically. To test the zero-shot capabilities of our method, we used data from completely different sources as the training set and test set in our experiments. Its superiority has been verified on publicly-available OCT datasets and the datasets collected via our home-built OCT systems. Along with the excellent generalization capability, our method also combines the high accuracy and real-time advantages of deep learning, which makes it promising to become a general-purpose analysis tool in the graphical user interface of OCT devices.