基于广义集成学习的超声图像甲状腺结节自动识别关键技术研究

With ultrasound examination as the commonly-used diagnostic approach, the inevitable subjective factors from the clinicians, such as clinical experience, knowledge level and working status, have varying influences on the clinical diagnosis of thyroid diseases. Thus, it urgently needs novel information technologies to improve the objectivity, high-efficiency and scientificity of thyroid nodule diagnosis. The most fundamental goals of this NSFC project are to systematically address a series of challenging theoretical issues, discover new knowledge through novel theoretic insights, and offer a suite of high-performance engineering solutions that are urgently needed in the unified local-global image feature representation, exemplar guided image ROI detection, the organic connection between low-level visual features and high-level semantics, and the dynamic integration of deep leaning and random forest regression. Our anticipated research outcomes include the automatic ROI discovery method based on lightweight lazy learning, the automatic thyroid nodule detection method based on hierarchical low-rank analysis over the discovered ROIs, the automatic thyroid nodule diagnosis based on dynamic ensemble learning, etc. Our research activities span from the novel theory and technology innovation to the algorithm and software development. Moreover, to fully evaluate and validate our new research results, we shall dedicate our engineering efforts to develop a prototype system, with clinical application to verify it. Key criteria to be utilized to judge this project's overall success include: the supported thyroid pathology number (at least 10 kinds), the accuracy in diagnosing the benign and malignant thyroid nodules (at least 90%), the accuracy in recognizing the thyroid patterns (at least 85%), runtime time performance for single ultrasound image processing (within 10 seconds), etc. The significant research and system outcomes of this research project will serve as critical technology supports to improve the diagnostic level of thyroid diseases, which will also greatly propel our nation’s high-quality medical resources to be more efficiently utilized.

目前以超声检查为主的甲状腺临床诊断会不同程度地受到医生临床经验、知识水平和工作状态等主观因素的影响，亟需借助信息技术来提高甲状腺结节诊断的客观性、高效性、科学性。项目通过医工结合的方式，以图像局部和全局视觉特征统一建模表示、样例驱动的兴趣区域识别检测、底层特征与高层语义有机关联、深度学习与随机森林回归动态集成为技术切入点，研究基于轻量级Lazy Learning的兴趣区域自动解析、基于兴趣区域层次化低秩结构分析的结节区域自动检测、基于动态集成学习的自动识别等关键科学问题，建立基于广义集成学习的超声图像甲状腺结节自动识别相关的理论和技术体系，并研发原型系统、开展临床验证应用。预期可实现支持不低于10种结节类型，结节良恶性判别准确率不低于90%、结节类型判别准确率不低于85%、单张图像判别时间不超过10秒的原型系统，为提高我国甲状腺疾病诊断水平、提升我国优势医疗资源利用效率提供信息技术支撑。

本项目通过医工结合的方式，以图像局部和全局视觉特征统一建模表示、样例驱动的兴趣区域识别检测、底层特征与高层语义有机关联、深度学习与随机森林回归动态集成为技术切入点，研究了基于轻量级Lazy Learning的兴趣区域自动解析、基于兴趣区域层次化低秩结构分析的结节区域自动检测、基于动态集成学习的自动识别等关键科学问题，建立了基于广义集成学习的超声图像甲状腺结节自动识别相关的理论和技术体系。相关成果在本领域顶级期刊和会议上发表高水平论文53篇，包括： SCI论文36篇，SCI论文占发表论文的67%；Q1分区论文19篇，占发表论文总数的35%，Q2分区以上论文2篇，占发表论文总数的3%；会议论文17篇，其中一篇获Best Paper Award，CCF推荐的国际会议论文9篇；获授权国家授权专利3项、另获受理专利6项；获批软件著作权3项。同时，项目研发了甲状腺结节智能诊断系统，可支持10多种结节类型，结节良恶性判别准确率达到96%以上、结节类型判别准确率达90%以上、单张图像判别时间在3秒以内的原型系统。研发的原型系统已在北京协和医院和中日友好医院等成功开展临床验证，为提高我国甲状腺疾病诊断水平、提升我国优势医疗资源利用效率提供了新人工智能技术支撑手段。