SFP-TFF: A Spatial-Frequency-Phase Tensor Fusion Framework for Semi-Supervised Medical Image Segmentation

Accurate medical image segmentation under limited annotation remains a critical challenge due to the high cost of expert labeling and the presence of low contrast, ambiguous boundaries, and imaging noise. While spatial domain features extracted by convolutional neural networks are effective for capturing local texture and edge details, they often suffer from instability in low-data or noisy scenarios. In contrast, frequency domain features exhibit stronger global regularity and robustness against noise, yet tend to overlook fine-grained structures. Motivated by the distinct distributional differences and complementary characteristics of spatial and frequency domain features, we propose SFP-TFF, a novel Spatial-Frequency-Phase Tensor Fusion Framework for semi-supervised medical image segmentation. Our approach adopts a dual-branch architecture that extracts both spatial features and frequency-phase representations, and fuses them through a tensor fusion module to align heterogeneous feature distributions and enhance feature complementarity. Furthermore, a semi-supervised learning strategy is employed to effectively leverage unlabeled data, improving generalization in data-scarce environments. Extensive experiments on benchmark medical image datasets demonstrate that SFP-TFF consistently outperforms existing state-of-the-art methods, particularly in low-label settings and challenging imaging conditions.