FDCF: Frequency Domain-Based Complementary Feature Fusion Network for Medical Image Segmentation

The complexity of anatomical structures and the frequent scarcity of annotated data make medical image segmentation a persistently challenging task. Most existing methods operate in the spatial domain, making them susceptible to indistinguishable noise and insufficiently sensitive to complex organ structures, which results in the degradation of subtle structural information. In response to these constraints, we present a Frequency Domain Complementary Feature Fusion Network (FDCF), which integrates two novel modules: the Hybrid Data Augmentation (HDA) module and the Bidirectional Frequency Domain Feature Fusion (BFDF) module. The HDA module enhances data diversity by linearly mixing two selected images and applying sequential geometric augmentations, effectively mitigating overfitting caused by limited training samples. The BFDF module fuses complementary features from dual branches in the frequency domain, thereby reducing noise interference while preserving detailed information. Validation across two open-access datasets, Synapse and RIM-ONE-R3, demonstrate that the proposed method significantly improves segmentation accuracy, highlighting its potential for clinical decision support in complex scenarios.