.png){kind=icon}
.png){kind=icon}
Return
WAAF-Net: Wavelet-guided asymmetric attention fusion network for dual-modal carotid plaque segmentation
.png){kind=icon}
PRE
AI
H
M
Y
C
E
Z
DOI：10.1016/j.eswa.2026.133855.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Existing multi-modality segmentation networks tend to aggregate heterogeneous features without discrimination, thereby overlooking the intrinsic discrepancy between ultrasound (US) and contrast-enhanced ultrasound (CEUS) in anatomical boundary representation and functional perfusion response. Even recent asymmetric architectures still mainly rely on spatial-domain fusion, which is vulnerable to ultrasound speckle noise and low-contrast boundaries, leading to insufficient cross-modal alignment and inaccurate plaque delineation. To alleviate these limitations, this paper proposes WAAF-Net, a wavelet-guided asymmetric attention fusion network tailored for dual-modal carotid plaque segmentation. Specifically, WAAF-Net adopts an asymmetric two-stream backbone to accommodate the modality-specific characteristics of US and CEUS, where the CNN branch focuses on local anatomical boundaries and textural patterns from US images, while the Transformer branch models CEUS-related cross-regional perfusion semantics and global contextual dependencies. Then, the Hybrid Wavelet Enhancement Block (HWEB) and Wavelet Fusion Enhancement Module (WFEM) extend cross-modal interaction from the spatial domain to the frequency domain. By explicitly decomposing features into low-frequency structural components and high-frequency boundary details, they enhance modality-specific representations while suppressing noise interference. Subsequently, the Shared-Projection Affinity Fusion Block (SPAFB) maps heterogeneous features into a common latent space, where affinity relationship modeling is performed to align and enhance modality-shared responses while compensating for modality-discrepant regions. Finally, the Prediction-Guided Feature Enhancement Block (PGFEB) introduces intermediate prediction maps as regional guidance to progressively aggregate multi-scale heterogeneous features, thereby recovering complex plaque morphology and weak boundary details. These three core modules follow a progressive design principle from modality-specific enhancement, to shared feature alignment, and further to multi-stage feature aggregation, forming a tightly coupled dual-modal segmentation framework. Experiments on a clinical US/CEUS carotid plaque dataset and the public CHAOS dataset demonstrate that WAAF-Net achieves superior segmentation performance and more robust generalization compared with state-of-the-art multi-modality baselines.
Keywords:
Asymmetric network
Carotid plaque segmentation
Dual-modal segmentation
Wavelet transform
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
Cited Papers