Motivation: Acquiring labeled data for fetal brain tissue segmentation is challenging and costly, limiting traditional supervised learning. Goal(s): This study aims to develop a semi-supervised method to enhance segmentation accuracy by addressing the difficulty in obtaining labeled data. Approach: Employing a single encoder and dual decoder structure, this method integrated a diffusion model to capture invariant features and refined precise features through guided consistent blocks. Results: Experimental outcomes demonstrated that this approach achieved high-precision image segmentation using a limited number of labeled samples, significantly enhancing accuracy while reducing reliance on expert input. Impact: This study introduces a semi-supervised fetal brain tissue segmentation method leveraging the diffusion model and guided consistency. It achieves comparable performance with fewer labeled samples, reducing manual marking time and advancing fetal brain diagnosis.
Qi et al. (Tue,) studied this question.