基于深度学习算法的甲状腺结节及颈部淋巴结超声管理系统性研究

近几十年全球甲状腺癌发病率持续上升，其淋巴结转移率高达20%-50%。甲状腺癌影像诊断首选超声，在我国医疗资源不均衡，影像诊断任务繁重复杂，亟需一种根据影像自动化、快速准确诊断的方法。课题组前期构建了用于识别甲状腺癌的深度卷积神经网络模型，申请人作为共同第一作者在Lancet Oncology发表封面文章。本项目将在此基础上扩展数据，建立人工智能的数据清洗及标注方案，构建大规模甲状腺超声数据集，采用深度学习方法分析甲状腺结节及颈部淋巴结超声图像，开发包括甲状腺影像报告和数据系统、甲状腺癌淋巴结转移风险评估和颈部淋巴结定性及定位诊断3个子系统的人工智能管理系统，在多中心数据中与人类超声专家的诊断水平进行比较，并初步探索模型可解释性。大规模数据集的建立可为甲状腺超声人工智能研究提供基础，甲状腺结节及颈部淋巴结超声管理人工智能系统的开发，有望改善部分机构诊疗水平，为临床决策提供支持。

The incidence of thyroid cancer has continued to rise worldwide over the past few decades, with lymph nodes metastasis rate as high as 20%-50%. Ultrasound is the first choice for thyroid cancer imaging diagnosis.The medical resources in China are imbalanced and the tasks of imaging diagnosis are overload and complicated，then an automated，rapid and accurate diagnostic method developed for thyroid imaging is urgently needed.The research group constructed a deep convolutional neural network model for identifying thyroid cancer. The applicant, as co-first author, published a cover article on Lancet Oncology for the study. In this study, we will expand the data set and establish a scheme to clean and label database by artificial intelligence to set up a large-scale thyroid ultrasound imaging data set. We will use deep learning algorithms to analyze thyroid nodules and cervical lymph nodes ultrasound images and develop an Artificial Intelligence Management System (AIMS) including three subsystems: thyroid imaging reporting and data system, thyroid cancer progress to lymph nodes metastasis assessment and the lymph nodes qualitative and location diagnosis. The AIMS will be compared with human ultrasound specialist on diagnostic levels in multicenter. Further, we will explore the model interpretability. This large-scale data set can provide the basis for the development of thyroid ultrasound artificial intelligence. The development of AIMS for ultrasound management of thyroid nodules and cervical lymph nodes will improve the level of diagnosis and treatment in some institutions and provide support for clinical decision-making in thyroid cancer.