基于全尺度网络深度度量学习与图像深度推衍的失智症辅助诊断研究

The dementia disease is widely acknowledged as one of the most serious diseases for elder populations in China at the current stage. Also, it is easily perceived that, timely and accurate diagnosis of the dementia disease is crucial for its later treatment. In this project, arterial spin labeling magnetic resonance images-based dementia disease diagnosis will be comprehensively and deeply investigated with the help of up-to-dates deep learning techniques. There are several key problems to be studied, which include 1) the deep derivation of arterial spin labeling images from structural magnetic resonance images; 2) single-pixel-based partial volume effects correction from time and deep latent domains; 3) deep metric learning based on full-scale deep residual networks. It is necessary to point out that, a new arterial spin labeling images database will be constructed using images both collected from actually-scanned patients and deep derivations from the well-known ADNI-1 database. A series of novel deep learning models and algorithms will be proposed from this project, and the diagnosis accuracy of this newly constructed database is expected to be no less than 95% after carrying out all studies in this project. To sum up, this project is an interdisciplinary one between machine learning and medical image processing. All three key problems in this project are first raised and studied, which strongly indicate their novelties. It is also important to mention that, the research aim of this project complies prefectly well with that of the current 3-year AI development scope in China, which was proposed by MIIT of China in December 2017.

失智症是已进入人口老龄化社会的我国愈发常见的老年人疾病，尽早且精确地确诊失智症病情对后期缓解甚至抑制患者病情意义重大。本项目拟采用前沿深度学习技术，对基于动脉自旋标记(ASL)图像实现失智症辅助诊断过程中一系列关键科学技术问题展开深入研究。内容包括：1）ASL图像的深度异构推衍；2）基于时域/深度变化域的ASL图像单像素点部分容积效应改善；3）基于全尺度深度残差网络的深度度量学习。本项目拟组建一个面向1200位失智症患者的ASL图像新数据库，其中包括自行采集数据和国际ADNI-1数据库中深度推衍数据两部分。本项目拟通过三个关键科学技术问题的研究,提出一系列深度学习前沿模型和创新算法，实现在新数据库中各类失智症病情检出率95%以上的目标。本项目属机器学习与医学图像处理领域交叉研究，三个关键科学技术问题的提出和研究在国内外属首次，创新性显著。本项目研究目标与工信部AI三年规划目标一致。

失智症是已进入人口老龄化社会的我国愈发常见的老年人疾病，尽早且精确地确诊失智症病情对后期缓解甚至抑制患者病情意义重大。本项目采用了前沿深度学习技术，对基于动脉自旋标记图像实现失智症计算机辅助诊断过程中一系列关键科学技术问题展开了具体且深入的研究。主要研究内容包括：1)动脉自旋标记图像的深度异构推衍，2)基于时域/深度变化域的动脉自旋标记图像单像素点部分容积效应改善，3)基于全尺度深度残差网络和深度度量学习技术的失智症计算机辅助诊断等工作。本项目组建了一个具备动脉自旋标记图像、结构性磁共振图像等模态在内的失智症患者图像数据库，对现有的国际上通行的ADNI数据库形成了良好补充。本项目主要研究成果通过34篇国际知名学术期刊或会议的论文形式发表，其中包括了IEEE Transactions on Medical Imaging, IEEE Transactions on Multimedia, Information Sciences, Pattern Recognition, Neurocomputing, MICCAI, ACM-Multimedia等高质量学术期刊和会议论文。本项目相关成果还包括江西省自然科学二等奖，国际学术会议ICITBE最佳论文奖、国际学术会议ICCEAI最佳论文奖等国内外重要学术奖项。培养博士研究生3名、硕士研究生11名；圆满完成了项目申请书中计划的各项工作任务。