SIGMA: spectral-implicit guidance for Multi-Teacher adaptation in unsupervised cross-modality 3D medical image segmentation

Unsupervised domain adaptation (UDA) for three-dimensional (3D) cross-modality medical image segmentation is affected by domain shifts arising from heterogeneous imaging physics and acquisition protocols. Existing methods remain limited by ringing artifacts from hard spectral truncation, unreliable pseudo-label supervision, and fixed single-pass inference. This paper presents Spectral-Implicit Guidance for Multi-Teacher Adaptation (SIGMA), a unified integration framework composed of Ringing-suppressed Soft Spectral Alignment (RSSA), Reliability-aware Multi-Teacher Fusion (RMTF), and Cognitive Implicit Refinement Adapter (CIRA). RSSA performs smooth low-frequency appearance transfer while preserving high-frequency anatomy; RMTF generates uncertainty-aware soft pseudo-labels from heterogeneous teacher disagreement; and CIRA conducts lightweight inference-time refinement over high-risk regions with the segmentation backbone frozen. Experiments on Brain Tumor Segmentation 2021 to Brain Tumor Segmentation 2023 Sub-Saharan Africa (BraTS2021-to-BraTS2023 SSA) and Multi-Modality Whole Heart Segmentation 2017 (MM-WHS 2017) magnetic resonance imaging (MRI)-to-computed tomography (CT) show that SIGMA achieves the highest average mean Dice among the evaluated methods and competitive boundary/volume performance. For MM-WHS, the observed differences are interpreted as numerical trends because the paired tests do not reach statistical significance with the limited evaluation cohort. The results suggest that spectral guidance, reliability-aware multi-teacher supervision, and localized implicit refinement provide complementary functions within a unified cross-modality 3D adaptation pipeline.