Elastic ultrasound image segmentation of thyroid nodules based on transfer learning

Thyroid nodule elastography ultrasound images contain information about nodule stiffness and morphology. Accurate segmentation of these images can significantly enhance the accuracy of thyroid nodule diagnosis. While deep learning-based thyroid nodule ultrasound image segmentation has shown promising results, its accuracy in elastography image segmentation remains limited due to the small dataset sizes. To address the low segmentation accuracy of thyroid nodule elastography images, we propose a transfer learning-based method that leverages shared features between grayscale ultrasound and elastography images. We first introduce a large-kernel attention mechanism to develop a multi-scale feature extraction network, capturing both local structural details and long-range dependencies in thyroid nodules. U-Net is then employed as the backbone network to build the grayscale ultrasound segmentation model. On this basis, we use the conditional generative adversarial network (CGAN) to align the feature distributions of grayscale ultrasound images (the source domain) and elastography ultrasound images (the target domain). A weight-sharing strategy is applied to transfer the feature extraction parameters, establishing a segmentation model for elastography ultrasound images. Experimental results show that the proposed segmentation method affords Dice coefficient, intersection over union (IOU), Recall, and Precision values of 77.09%, 65.20%, 78.19%, and 81.05%, respectively.