Dual-stream Assisted U-Net for Thyroid Nodule Segmentation
Qi Zhou, Hang Yang, Chuangeng Tian, Yu Hui, Lu Tang · 2023
Abstract—Accurate and effective segmentation of thyroid nodules in ultrasound images is critical for computer-aided nodule diagnosis and treatment. However, due to the ill-defined margin of ultrasound images and the complexity of thyroid tissue structure, accurate segmentation of fine contours is still challenging. We propose a dual-stream network-assisted U-Net (DA-UNet) for these issues, which fully integrates features extracted by convolutional neural networks and vision transformer (ViT), applying them to the decoder and skip connection stages. The model is evaluated on the publicly available dataset TN3K, and the experimental results show that DA-UNet achieves 84.43% and 85.55% in Dice coefficient and F1-score respectively, with an average improvement of 3.75% and 3.12% compared to the current popular segmentation networks, reaching the optimal level.