Intelligent Segmentation of Thyroid Nodule Ultrasound Images Based on Benign--Malignant-Aware Semantic Prototype Calibration and Class-Conditional Boundary Refinement

Abstract

Thyroid nodule ultrasound images often suffer from blurred boundaries, complex echo noise, and subtle morphological differences between benign and malignant lesions, which bring challenges to automatic segmentation and assisted diagnosis. To address these issues, this paper proposes a benign--malignant-aware segmentation framework integrating adaptive semantic prototype calibration and class-conditional boundary refinement. The proposed method adopts SegFormer as the backbone network and formulates thyroid nodule segmentation as a three-class task involving background, benign, and malignant regions. Specifically, the ASPC module enhances the discriminative ability between benign and malignant regions by learning category-level semantic prototypes, while the CCBR module adaptively corrects ambiguous contours by combining predictive entropy and foreground boundary responses. Experiments are conducted on the public TN3K dataset and an external clinical validation dataset. The results show that the proposed method achieves PA, mIoU, mDice, and mPrecision values of 0.9449, 0.6427, 0.7517, and $0.8147$ on the TN3K dataset, and 0.8521, 0.6124, 0.7328, and 0.7996 on the external validation dataset, respectively. Ablation experiments and Grad-CAM visualization further verify the effectiveness of the proposed modules in semantic discrimination, boundary refinement, and model interpretability.