Contrast-Enhanced VM-UNet for Thyroid Nodule Segmentation Based on Wavelet Features and Dual-Branch Edge Learning

Accurate segmentation of thyroid nodules from ultrasound images is critical for clinical diagnosis and treatment, as manual analysis is labor-intensive and subject to inter-observer variability. However, existing models often struggle with edge detail capture and discriminative representation of nodule regions, limiting their clinical utility. To address these challenges, we propose a contrast-enhanced VM-UNet for thyroid nodule segmentation, which integrates three key innovations: (1) Wavelet transform to extract frequency-domain features, compensating for VM-UNet's limitations in capturing edge details using only spatial-domain features, (2) a dual-branch structure (full mask and edge mask) in the upsampling stage to strengthen edge feature learning, (3) a contrastive loss mechanism between background and target regions to enhance the model's discriminative representation of nodules. Quantitative experiments demonstrate that our method achieves a Dice similarity coefficient (DSC) of 93.61% and a mean Intersection over Union (mIoU) of 88.56% on the TN3k dataset. Crucially, on an external clinical dataset of 360 cases, our model maintained high performance (DSC, 92.87%), significantly outperforming the baseline VM-UNet. Our proposed approach effectively combines wavelet transform, dual-branch edge learning, and contrastive learning, providing a robust technical solution for accurate thyroid nodule segmentation and supporting more reliable clinical decision-making.