Pyramid Shape-Aware Semi-supervised Learning for Thyroid Nodules Segmentation in Ultrasound Images

The accurate segmentation of thyroid nodules in ultrasound (US) images is critical for computer-aided diagnosis of thyroid cancer. While the fully supervised methods achieve high accuracy, they require a significant amount of annotated data for training, which is both costly and time-consuming. Semi-supervised learning can address this challenge by using a limited amount of labeled data in combination with a large amount of unlabeled data. However, the existing semi-supervised segmentation approaches often fail to account for both geometric shape constraints and scale differences of objects. To address this issue, in this paper we propose a novel Pyramid Shape-aware Semi-supervised Learning (PSSSL) framework for thyroid nodules segmentation in US images, which employs a dual-task pyramid prediction network to jointly predict the Segmentation Maps (SEG) and Signed Distance Maps (SDM) of objects at different scales. Pyramid feature prediction enables better adaptation to differences in nodule size, while the SDM provides a representation that encodes richer shape features of the target. PSSSL learns from the labeled data by minimizing the discrepancy between the prediction and the ground-truth and learns from unlabeled data by minimizing the discrepancy between the predictions at different scales and the average prediction. To achieve reliable and robust segmentation, two uncertainty estimation modules are designed to emphasize reliable predictions while ignoring unreliable predictions from unlabeled data. The proposed PSSSL framework achieves superior performance in both quantitative and qualitative evaluations on the DDTI and TN3k datasets to State-Of-The-Art semi-supervised approaches. The code is available at https://github.com/wuliZN2020/Thyroid-Segmentation-PSSSL .