High frequency edge network for accurate cardiac structure segmentation

Accurate cardiac structure segmentation is intrinsically a boundary delineation problem, where discriminative anatomical cues are largely encoded in high frequency components. We develop High Frequency Edge Network (HF-EdgeNet), a high frequency driven encoder decoder framework that incorporates structural cues throughout cardiac magnetic resonance imaging (MRI) segmentation. Specifically, HF-EdgeNet uses the High Frequency Edge Transformer (HF-EdgeT) to inject high frequency guidance into self-attention, introduces the High Frequency Adaptive module (HF-Adapte) to compensate for high frequency degradation during down sampling, and designs the Semantic Edge Bridge (SEB) block for high frequency semantic re-alignment during decoding. Experiments on ACDC and M&Ms show improved average dice scores and favorable boundary related performance over strong segmentation baselines. These results support explicit high frequency modeling for boundary sensitive cardiac image segmentation.