A key instance-guided frame-to-video information fusion network for thyroid ultrasound video instance segmentation

Video instance segmentation (VIS) in thyroid ultrasound (US) images has the advantages of automatically segmenting nodules and assigning instance labels, which aids in counting nodules and determining the thyroid key plane. However, thyroid US videos are characterized by abundant background noise and echo similarity. Coupled with the continuous variations in the shape of tissue structures along the video stream, it becomes extremely challenging for current mainstream VIS methods to identify the temporal dynamic change features, limiting their performance in segmentation and tracking. To address this issue, this paper proposes a key instance-guided frame-to-video information fusion network (NoVIS) for thyroid US video instance segmentation. First, a frame-level instance information enhancement module is introduced to capture key instance information, constructing robust key instance features and enhancing similarity within the same instance. Second, a key instance-guided global information aggregation module is proposed. This module leverages key instance features to continuously update the global temporal information, aiming to capture the dynamic variation characteristics of structures along the video stream. Finally, a matching mechanism ensures consistency between training and inference, further improving tracking stability. Extensive experiments were conducted on a collected thyroid US video dataset, which demonstrated that NoVIS significantly outperforms other methods in both tracking and segmentation. Specifically, NoVIS achieves an AP of 58.09% and an AP₅₀ of 81.33%, which are 3.37% and 5.09% higher than those of the baseline method, respectively. Furthermore, the ability of the proposed method to determine the thyroid key plane and count nodules is demonstrated, which is highly valuable for fully automated thyroid clinical diagnosis.