A semi-supervised segmentation method and system for pulmonary bronchial tree based on enhanced pseudo-label mechanism

Abstract

This invention provides a semi-supervised segmentation method and system for the tracheal tree based on an enhanced pseudo-labeling mechanism. The method includes: acquiring a tracheal tree labeled image; obtaining a tracheal tree pseudo-label image through a semi-supervised learning algorithm; optimizing the tracheal tree pseudo-label image using an enhancement network to obtain intermediate predicted pseudo-label images; processing the intermediate predicted pseudo-label images using a maximum connected component algorithm to obtain a first feature pseudo-label image and performing post-processing; extracting a second feature branch image based on the processed first feature pseudo-label image and the tracheal tree labeled image; obtaining a second feature pseudo-label image based on the second feature branch image and the second feature label through a semi-supervised learning algorithm; and synthesizing the intermediate predicted pseudo-label image, the first feature pseudo-label image, and the second feature pseudo-label image to generate an enhanced pseudo-label image. This method achieves model performance similar to a fully supervised model with a lower number of supervised annotations.

Images