Abstract
Retinal diseases, including myopic and diabetic retinopathies, require early detection through precise retinal-layer segmentation in optical coherence tomography images. Existing deep-learning models generalize poorly across devices (domain shifts, noise, and high annotation costs). We propose dual-level pseudo-label learning for segmentation (DPLSeg), an unsupervised segmentation model with a dual-level pseudo-label learning strategy and a hierarchical transformer encoder to enhance feature representation and domain adaptability. Validated on 850 optical coherence tomography images from three devices, DPLSeg achieves a mean intersection over union of 79.9%, surpassing DeepLab (75.2%) and DAFormer, reducing annotation needs by 80% and providing a scalable clinical diagnostic tool.