Few‐shot segmentation for esophageal OCT images based on self‐supervised vision transformer

Abstract

Automatic segmentation of layered tissue is the key to optical coherence tomography (OCT) image analysis for esophagus. While deep learning technology offers promising solutions to this problem, the requirement for large numbers of annotated samples often poses a significant obstacle, as it is both expensive and challenging to obtain. With this in mind, we introduced a self‐supervised segmentation framework for esophageal OCT images. In particular, the proposed method employs a masked autoencoder (MAE) for self‐supervised training and constructs the segmentation network by integrating a pretrained vision transformer (ViT) encoder with an attentive transformer decoder. In this case, the segmentation network has the potential to accomplish the few‐shot, or the more aggressive one‐shot segmentation, and achieve high‐quality segmentation performance. Experimental results on both a self‐collected mouse esophageal dataset and a public human esophageal OCT dataset confirm the advantages and practical significance of the proposed method.