基于多维融合信息的肺结节检测与良恶性鉴别

Lung cancer is a leading cause of deaths from malignant tumors in China, accounting for 22.5% of all cancer deaths. Here, the lung cancer screening for asymptomatic population is an effective way to achieve early detection, early diagnosis and early treatment of lung cancer. The cancer mortality rate can be reduced of 30% approximately if the cases are detected and treated early and the patient five-year survival rate can be improved obviously. The key technologies of lung nodules detection and lung nodules benign and malignant discrimination are researched in lung cancer screening process in this project. It is a great significance. The project proposed a two-steps lung cancer screening strategy. First, the low radiation DR/multi-projection image sequences are used for preliminary screening of asymptomatic population, and then the low-dose pulmonary CT images are used for review screening of suspected patients. The focus of the project include: 1) Lung nodule detection based on DR / multi-projection image sequences; 2) Lung nodule detection for low-dose CT image based on the multi-dimensional fusion information; 3) The method for pulmonary nodule benign and malignant discrimination; and 4) The database construction of the confirmed cases. The project will acquire the pulmonary nodules benign and malignant identify index for computer aided diagnosis in clinic, based on accurate extraction of lung nodules, analysis the imaging features of candidate lung nodules and constructed the pulmonary nodules benign and malignant discrimination model with multi-classifier. This study not only can obtain theoretical results in the field of medical image processing and analysis, but also can promote the launching of lung cancer screening in our country and reduce the social harm caused by lung cancer.

肺癌是我国死亡人数最多的恶性肿瘤，约占全部恶性肿瘤死亡人数的22.5%。对无症状人群进行肺癌筛查，是实现肺癌早发现、早诊断和早治疗的关键，可降低约30%的死亡率，提高患者术后五年生存期。本项目针对肺癌筛查过程中肺结节检测和良恶性鉴别等关键技术展开研究，具有重要意义。 项目提出先使用低辐射的DR/多投影相关序列影像对无症状人群进行初步筛查，再使用低剂量CT影像对疑似患者进行复查的肺癌筛查策略，重点研究了基于DR/多投影相关序列影像的肺结节检测、基于多维融合信息的低剂量CT影像中肺结节检测、肺结节良恶性鉴别及确诊病例数据库构建等问题。在实现肺结节准确提取、候选肺结节影像学征象分析及基于多分类器的肺结节良恶性鉴别模型构建的基础上，对检测到的肺结节给出良恶性综合鉴别指标，用于临床辅助诊断。本项研究不仅可在医学影像处理和分析领域取得理论成果，还会推动我国肺癌筛查工作的开展，降低肺癌所造成的社会危害。

本项目完成了研究计划内容，针对所提出的二级肺癌筛查策略，结合所构建的确诊病例样本库，研究具有鲁棒性的肺结节检测和良恶性鉴别关键技术，为解决临床肺癌筛查任务，提供良好的技术理论基础和实际问题解决方案。（1）构建了包括1200例病例的基于病理及PET影像诊断确定的良恶性肺结节病例样本库，可用作金标准对相关算法进行验证；（2）提出了CT影像中肺结节多种影像学特征建模方法，实现了影像学特征的定量描述，可将其用于对肺结节良恶性的判别；（3）基于深度学习的医学影像分析理论及其应用研究，构建多种网络模型，用于对肺结节初始检测、假阳性去除、良恶性判别等环节，获得了更好的分类和预测结果；（4）ELM（极限学习机）基础理论及其应用研究，解决了极限学习机输出不稳定问题，将其用于肺结节检测和分类；（5）提出基于多投影图像配准的肺结节检测算法，利用同一病人多投影图像中相近位置的候选结节互信息的配准算法，由此来减少检测结果中假阳性结节的数目；（6）提出多种目标检测算法用于肺结节的检测，采用了粒子滤波理论、Adaboost算法等技术理论基础；（7）提出了基于特征图像及灰度与形状模型的DR图像肺分割方法和基于支持向量机的DR图像肋骨分割方法，用于实现对影像的预处理，有效降低了检测的搜索空间和提高了检测结果的准确率。（8）提出了基于多维图像信息的肺结节良恶性鉴别方法，该方法基于二、三维肺结节影像学征象，采用模糊C均值分类器实现了肺结节的良恶性鉴别；（9）共发表学术论文18篇，其中SCI收录1篇，EI收录14篇，申请发明专利11项，获授权发明专利6项，依托本项目共培养硕士研究生16人，已毕业14人，其中2人获国家奖学金和校级优秀硕士学位论文。