AMSeg: A Novel Adversarial Architecture Based Multi-Scale Fusion Framework for Thyroid Nodule Segmentation

The thyroid gland is an important and essential endocrine organ for the human body that regulates metabolism, growth and development by secreting thyroid hormones. Thyroid nodules are irregular masses caused by lesions that can reflect the main clinical manifestation of thyroid abnormalities. Delineating the boundaries of thyroid nodules from ultrasound images is an indispensable part of computer-aided diagnosis systems and medical imaging diagnosis for thyroid diseases. However, automatic segmentation of thyroid nodules is still a challenging task because of the similarity between the heterogeneous appearance and its background. In this study, we propose a novel framework for thyroid nodule segmentation that exploits multiscale anatomical features to build a late-stage fusion method based on adversarial training. The introduced architecture can handle blurred and uneven tissue regions during the thyroid nodule segmentation process. By adversarial training approaches, our segmentation block S adopts a framework with three different fusion scales, and discrimination block D employs a fully convolutional encoder-decoder architecture. Extensive experimental results demonstrate that AMSeg outperforms popular methods in terms of thyroid nodule segmentation. The proposed framework achieves Dice, Hd95 (95% Hausdorff distance), Jaccard, and precision values of 83.06%, 23.13, 74.18%, and 87.82%, respectively. As an end-to-end network, the AMSeg can effectively replace manual segmentation methods and has great prospects in clinical applications.