TEM-Net: A Tri-Channel Edge-Aware Multi-Scale Network for Thyroid Nodule Segmentation in Ultrasound Images

With the increasing detection of thyroid nodules in ultrasound screening, accurate nodule segmentation has become important for computer-aided assessment of clinically relevant features, such as contour regularity, aspect ratio, and margin sharpness. Although ultrasound is widely used as a first-line imaging modality in clinical practice, thyroid nodule segmentation remains challenging because of low tissue contrast, speckle-blurred boundaries, and large variations in nodule size and morphology. To address these challenges, we propose TEM-Net, a Tri-Channel Edge-Aware Multi-Scale Network for thyroid nodule segmentation. TEM-Net constructs a tri-channel representation from the raw ultrasound image, including the original grayscale image, a contrast-enhanced image, and a gradient-magnitude map, to highlight weak intensity differences and boundary-related cues in low-contrast ultrasound images. An Edge-Guided Feature Amplification (EGFA) module is introduced before the first down-sampling operation to emphasize boundary responses before spatial resolution is reduced. In addition, a Multi-Focus Cross-Scale Attention Refinement (MF-CAR) module is embedded into skip connections, combining dilated depth-wise convolutions with channel-spatial attention to improve the fusion of local boundary details and broader contextual information. Across three random seeds, TEM-Net achieves mean Dice scores of 0.8822 and 0.9066 and mean IoU scores of 0.7893 and 0.8291 on TN3K and DDTI, respectively, showing competitive performance compared with representative segmentation methods.