HDDI-Net: Hierarchical dual-domain interaction network for robust and efficient ultrasound lesion segmentation.
Authors
Affiliations (4)
Affiliations (4)
- School of Computer Science and Technology, Southwest University of Science and Technology, Mianyang, 621010, China. [email protected].
- School of Computer Science and Technology, Southwest University of Science and Technology, Mianyang, 621010, China.
- School of Artificial Intelligence, Southwest University, Chongqing, 400715, China.
- School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, 100876, China.
Abstract
Lesion segmentation in B-mode ultrasound remains a critical challenge due to speckle interference, low tissue contrast, and constrained computational budgets in portable devices. Accurate delineation of lesions is essential for diagnosis, treatment planning, and longitudinal monitoring. Existing lightweight spatial models often fail to suppress speckle-dominated noise, whereas transformer-based methods exceed point-of-care computational limits. To address these challenges, we propose HDDI-Net, a Hierarchical Dual-Domain Interaction Network integrated into a macro-micro region-of-interest pipeline. HDDI-Net combines multi-kernel spatial feature extraction with discrete cosine transform (DCT)-guided frequency gating to effectively suppress speckle while preserving morphological integrity. At the network bottleneck, a Prototype-Guided Semantic Consistency (PGSC) module approximates global context without high computational cost. Extensive experiments on the BUSI and TN3K datasets, along with zero-shot cross-domain evaluation on UDIAT, demonstrate that HDDI-Net achieves improved accuracy-efficiency trade-offs compared with evaluated lightweight CNNs and transformer-based segmentation models, operating with only 1.7M parameters and a moderate computational footprint of 3.3 GFLOPs, achieving up to 5.0 percentage-point IoU improvement in zero-shot evaluation and up to 3.9 percentage-point Dice improvement over the strongest competing baseline. The proposed approach shows promising potential for resource-constrained ultrasound lesion segmentation and may support future integration into computer-aided diagnosis pipelines after further deployment-oriented optimization and clinical validation. To facilitate reproducibility, the source code and pre-trained models are publicly available at https://github.com/DPDP-root/HDDI-Net.