Learning OCT segmentation from a single label

Abstract

Deep learning boosts the performance of automatic OCT segmentation, which is a prerequisite for standardized diagnostic and therapeutic procedures. However, training deep neural network requires laborious data labeling, and the trained models only work well on data from the same manufacturer, imaging protocol, and region of interest. Here we propose a novel learning method to reduce labeling costs. By labeling and training on a single image, we achieved segmentation accuracy comparable to that of a U-Net model trained on ~25 to 50 labeled images. This reduction in labeling costs could significantly improve the flexibility and generalization of deep-learning-based OCT segmentation.