Answers are grounded in this paper and cite their sources.
Accurate thyroid nodule segmentation is important for ultrasound image analysis, yet low contrast, blurred boundaries, and heterogeneous background responses jointly affect lesion recognition, contour localization, and mask generation. More specifically, lesion structures and irrelevant high-frequency background patterns may produce similar local responses, deep and shallow features are prone to semantic-contour misalignment in weak-boundary regions, and classification, localization, and mask prediction may compete when they share the same features. To address these issues, this paper proposes EffiThy-Net, a lightweight lesion-calibrated and boundary-preserving segmentation network that establishes a coordinated processing pipeline for lesion-aware feature calibration, lesion-contour semantic alignment, and boundary-preserving prototype-decoupled mask decoding. First, the Lesion-Aware Alternating Calibration Block (LAC-Block) alternates feature propagation and response calibration to suppress irrelevant high-frequency background responses while preserving lesion-related structures. Second, the Lesion-Contour Semantic Alignment Module (LCSA) uses deep lesion semantics to constrain shallow contour details, thereby alleviating feature misalignment in weak-boundary regions. Third, the Boundary-Preserving Prototype Decoupling Head (BPD-Head) reduces interference among prediction branches through task decoupling and prototype-based mask reconstruction. Experiments on the TN3K and TNUI-2021 datasets are conducted under a unified pixel-level mask evaluation protocol. The results show that EffiThy-Net achieves a favorable balance between accuracy and efficiency under the current setting. Compared with DeepLabv3+, the proposed method improves DSC by 3.1 percentage points on TNUI-2021 while reducing the number of parameters by 92.5% and GFLOPs by 91.4%. Additional controlled additive Gaussian perturbation experiments show that, on TN3K, when the perturbation level reaches \(σ =20\) , the drops in mAP@50 and DSC for EffiThy-Net are 2.5 and 2.3 percentage points, respectively, which are smaller than those of the included YOLO baselines. These results indicate lower sensitivity to the evaluated additive Gaussian perturbations, but they do not establish robustness to physical multiplicative ultrasound speckle.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: