Automated segmentation of thyroid tissue and carotid artery in ultrasound videos using deep learning: A foundational step toward AI-assisted thyroid diagnostics

Background Ultrasound (US) is the primary modality for thyroid evaluation, yet its diagnostic accuracy remains operator dependent. Reliable automated segmentation of thyroid anatomy and adjacent vascular structures is a prerequisite for AI-based nodule detection, risk stratification and future robot-assisted scanning, supporting more standardized preoperative decision-making in endocrine surgery. Methods A retrospective dataset of 400 B-mode US videos from 380 patients was collected between 2016 and 2022. All patients were referred for symptomatic or incidentally identified thyroid nodules. A video-based nnU-Net deep learning model was trained to segment the thyroid gland and carotid artery. Data were split 70/15/15 into training, validation, and test sets. Performance was evaluated using the Dice similarity coefficient. Results The model achieved high segmentation performance, correctly delineating thyroid tissue with a Dice coefficient >0.91. Carotid artery segmentation reached a Dice coefficient >0.96, enabling highly accurate localisation of this key vascular landmark. Notably, 88% of the videos included thyroid nodules, both malignant and benign, indicating that the model was developed on a variety of clinical cases rather than homogeneous datasets. Conclusions This study demonstrates the feasibility of real-time, video-based segmentation of thyroid and carotid anatomy as a foundational step toward AI-assisted endocrine surgery. Beyond reducing operator dependency and enabling automated nodule assessment, the video-based approach supports 3D anatomical reconstruction, which may enhance future surgical planning, navigation and robotic guidance within thyroid surgery workflows.