FreqBlueprint: Frequency-Bank Guided Dynamic Graph Convolution for Medical Image Segmentation

In medical images, lesions are often obscured by complex background textures and ambiguous anatomical boundaries. To derive accurate segmentation boundaries from such challenging visual patterns, we move beyond the one-size-fits-all design of fixed convolutional kernels in conventional convolutional neural networks and propose a frequency-bankbased dual-domain graph convolutional architecture, termed FD-GCN. Inspired by the diagnostic workflow of pathologists, who first examine the overall tissue organization at low magnification and then inspect fine cellular morphology at high magnification, our model establishes an instructionguided modular assembly mechanism. Specifically, we design a Dynamic Graph Frequency Bank (DGFB) that adaptively generates spatial- and frequency-domain weights from global contextual information. These weights serve as guidance signals for Frequency-Guided Scale-Adaptive Grid Graph Convolution (FSGC) to dynamically assemble multi-scale graph neighborhoods, thereby enlarging the effective receptive field and enhancing structural modeling while maintaining moderate computational complexity. In addition, Spatial-Frequency Fusion Attention (SFFA) and Adaptive Cross-stage Feature Fusion (ACFF) are introduced to promote multi-scale dualdomain feature interaction and alleviate the semantic discrepancy between the encoder and decoder. FD-GCN achieves a Dice score of 85.72% on the Synapse multi-organ segmentation dataset and an IoU of 86.26% on the ISIC2016 dataset, outperforming existing state-of-the-art methods.