FetalSFUDA: Source-Free Unsupervised Domain Adaptation for Fetal Brain Extraction from Different Centers or MRI Sequences

Motivation: Brain extraction is fundamental in fetal brain MRI 3D reconstruction and analysis. However, the inter-domain generalization of pre-trained deep leaning models is poor since scanner and imaging parameters are diverse among datasets. Goal(s): Achieve robust fetal brain extraction on out-of-distribution (OOD) target domains using source-free unsupervised domain adaptation (SFUDA) method. Approach: Parameters of a pre-trained model are optimized for domain adaptation by minimizing a label-free entropy loss and incorporating a class-ratio prior constraint. Results: The pre-trained model, after applying SFUDA, demonstrates significantly enhanced fetal brain extraction performance on OOD target domains. Impact: Source-free unsupervised domain adaptation addresses the problem that the pre-trained fetal brain extraction model is inaccurate for data acquired with different scanning hardware and parameters. Moreover, our work supports cross-center fetal studies and promotes practical clinical diagnostic applications.