Thyroid nodule segmentation in ultrasound images using U-Net with ResNet encoder: achieving state-of-the-art performance on all public datasets

Ultrasound imaging plays a vital role in evaluating thyroid nodules, aiding in the assessment of malignancy risk, monitoring size progression, and serving as a guiding tool for thyroid nodule biopsies. Computer-aided diagnosis (CAD) systems have emerged to assist in diagnosing thyroid lesions, reducing unnecessary biopsies, and contributing to the overall improvement of diagnostic accuracy. The segmentation process plays a crucial role in CAD systems because it marks the region of interest. If segmentation were sufficiently accurate, then it would improve the entire diagnostic process and bring CAD systems closer to routine clinical practice. As far as we know, there are currently only three publicly available datasets of ultrasound images of the thyroid gland that can be used for the purpose of thyroid nodules segmentation. The Thyroid Digital Image Database (TDID) is a long-standing benchmark dataset but faces limitations due to the data ambiguities. The TN3K dataset is more robust than TDID, and the Thyroid Ultrasound Cine-clip dataset offers recent alternatives. In this paper, we implemented a deep learning segmentation model based on UNet with a ResNet encoder. We trained this model on all available data and evaluated it on the TN3K test set. The achieved results for the Dice score, IoU score, accuracy, precision, and recall were 84.24%, 75.48%, 97.24%, 82.75%, and 88.98%, respectively. These results represent the most advanced state-of-the-art scores compared to previously published studies and demonstrate that UNet with a ResNet encoder has the capability to accurately segment thyroid nodules in ultrasound images.