Unsupervised Domain Adaptation with SAM-RefiSeR for Enhanced Brain Tumor Segmentation

Authors:

Abstract
Robust brain-tumor segmentation in MRI must withstand domain shifts from heterogeneous scanners and protocols. Unsupervised domain adaptation (UDA) can exploit plentiful unlabeled data, yet many approaches erode boundaries, suppress tumor cues, or amplify errors from noisy pseudo-labels. Generalpurpose models like SAM also struggle on MRI due to the domain gap and a lack of morphology-aware consistency. We propose SAM-RefiSeR, a two-phase UDA framework that integrates SAM for reliable, annotation-efficient segmentation. Phase I reduces source-target discrepancy via Fourier-based frequency adaptation and adversarial feature alignment, aligning appearance and representations while preserving anatomy. Phase II adopts a student-teacher scheme in which SAM first refines pseudo-labels, then gates them with confidence- and morphology-aware criteria to suppress unreliable masks and curb error propagation. Across diverse cross-modality settings, SAM-RefiSeR consistently surpasses strong UDA baselines, particularly under severe shifts. By improving boundary fidelity and robustness without additional labels, SAM-RefiSeR brings brain-tumor segmentation closer to practical, generalizable clinical deployment.