Contrastive Mutual Learning: A Semi-Supervised Method for 3D Fetal Brain Segmentation

Motivation: Fetal MRI and accurate segmentation are essential for clinical use. Manual segmentation is time-consuming, and deep learning methods require substantial labeled data, which is scarce. Goal(s): This study aims to propose a semi-supervised framework for 3D fetal brain segmentation that achieves satisfactory segmentation results while significantly reducing the amount of labeled data required. Approach: A transformer-based network structure utilizing contrastive learning, mutual learning, and incorporating consistency loss serves as the semi-supervised learning framework. Results: Using eight labeled samples, the proposed method achieved a mean Dice score of 0.81 on the test set, surpassing conventional supervised methods that yielded Dice scores of 0.25-0.35. Impact: This study implemented a semi-supervised learning approach to address the challenges of limited labeled data and high annotation costs in 3D fetal brain segmentation. The performance demonstrates the proposed algorithm's robust potential.