Multi-annotation agreement and prediction consistency networks: Improving semi-supervised segmentation of medical images with ambiguous boundaries

Shuai Wang, Tengjin Weng, Jingyi Wang, Kai Zhao, Yang Shen, Zhidong Zhao, Yixiu Liu, Pengfei Jiao, Zhiming Cheng, Yaoqi Sun, Yaqi Wang

Abstract: Highlights•We combine multi-annotated and semi-supervised segmentation and propose the MSE-Nets, aiming to improve the performance of ambiguous boundaries medical image segmentation in scenarios with a small amount of multi-annotated data and a large number of unannotated.•We propose the NPCE module for separating pixel-level (dis)agreement information from multi-annotated data for two purposes: agreement information is directly input into the network as reliable prior knowledge and; disagreement information is replaced based on whether the prediction results are consistent for label refinement.•We propose the MNPS module use the predicted consistent masks of multiple networks as the ground truth for unannotated images. The MNPS serves two benefits: strengthening the consistency between networks by incorporating additional intrinsic image knowledge from a substantial volume of unannotated data, which can be transferred to enhance the prediction consistency of multi-annotated data, and; preemptively circumventing the adverse impact of imprecise pseudo-labels on the network’s learning.

External IDs:doi:10.1016/j.artmed.2025.103289