From static wavelets to learnable spectral modulation after wavelet decomposition: a guided region-contour consistency network for breast ultrasound segmentation

Abstract

Accurate medical image segmentation remains challenging due to ambiguous lesion boundaries, low contrast, and complex structural variability. Existing approaches that rely primarily on spatial-domain feature learning often struggle to capture fine-grained details and enforce consistency between region prediction and contour localization. In this paper, we propose MFCD-Net, a novel segmentation framework that jointly models structural semantics and boundary-aware information in a unified manner. Specifically, we introduce a Frequency-Decoupled Modulation (FDM) module, which applies a fixed wavelet decomposition to separate intermediate features into complementary components, then employs learnable spectral modulation to enhance boundary-relevant information. Unlike static frequency transforms, FDM enables task-driven modulation of boundary cues while preserving spatial context. Furthermore, we develop a Guided Region–Contour (GRC) module, where boundary-aware responses are utilized to guide region prediction via an adaptive gating mechanism, promoting alignment between object interiors and their corresponding contours. The decoder incorporates multi-scale context aggregation to effectively fuse hierarchical features and improve robustness across varying lesion sizes. Extensive experiments on multiple public datasets demonstrate that MFCD-Net consistently outperforms state-of-the-art methods in terms of Dice, IoU, and boundary accuracy, highlighting its effectiveness for precise and reliable medical image segmentation.