Fully automated thyroid ultrasound screening utilizing multi-modality image and anatomical prior

There is a high prevalence of thyroid nodules in the general population. Early detection is essential for the treatment of malignant thyroid nodules. Ultrasound has the advantage of being non-invasive and radiation-free, and is sufficient for early diagnosis of malignant nodules. However, ultrasound image acquisition relies on manual manipulation of the probe by the sonographer, making it difficult to screen populations for malignant thyroid nodules. Robotic ultrasound systems (RUS) are expected to replace sonographers in ultrasound scanning and improve the standardization and reproducibility of ultrasound examinations. However, there is currently no RUS that can fully autonomously screen for thyroid nodules. In this paper, we propose a fully automated two-step search method that mimics a physician's thyroid scanning protocol. First, using the body surface structure observed by the RGBD camera, we use a bimodal detection network for the initial localization of the human neck. The bimodal detection network achieves an average accuracy of 0.986. Second, we designed a tracking strategy based on the anatomical prior of the human neck to navigate the probe for thyroid ultrasound acquisition. An network-based visual servoing method enables shadow prevention and target tracking to be performed uniformly. In vitro experiments demonstrated the effectiveness of the protocol.