基于深度学习的心脏SPECT影像数据自动分析与量化算法

SPECT心脏成像检查对于心血管疾病的临床诊断、预后分析等有重要的价值，而快速、精准的心脏图像处理和量化分析结果将直接影响临床诊断的准确性。目前常用的分析软件存在三大技术瓶颈，分别是：将SPECT图像转向至临床标准视图需医师的复杂手动操作，且易引入人工误差；常规分割算法对低分辨率带运动伪影的心脏的分割精度不足；靶心图自动量化模型对数据库依赖较高，临床分析智能化与准确性欠缺。这些问题容易造成临床诊断的错判，制约了临床应用的发展。针对这类问题，本项目旨在设计一套基于深度学习的左心室图像自动旋转算法、精准三维图像自动分割神经网络算法和基于图卷积网络的靶心图自动量化模型，整合形成一款用以SPECT心脏图像自动处理及量化分析的新型系统，以提供更准确、更智能的左心室自动转向、定位、精准分割，及更全面且准确的靶心图自动量化分析结果，服务国内核医学临床应用，推进国内核医学心脏病检测及诊断的发展。

SPECT cardiac imaging is of great value in the clinical diagnosis and prognosis of cardiovascular diseases. Rapid and accurate cardiac image processing and quantitative analysis will directly affect the accuracy of clinical diagnosis. There are three major technical bottle-neck problems in the analysis software currently in use: The complex manual operations required to orientate the SPECT reconstructed images to standard clinical views may lead artificial errors; The low resolution of the SPECT cardiac images and the respiratory and heart beating motion lead to the lack of accuracy of traditional segmentation methods; The automated bullseye quantitative software relies heavily on the accurate database and lacks intelligence and accuracy in analysis. These problems can lead to incorrect results in clinical diagnosis and limit the development of clinical applications. To address these problems, this project aims to design a novel deep-learning based system for automatic processing and quantitative analysis of cardiac SPECT images, which include a left ventricle images auto-reorientating algorithm, an accurate 3D image auto-segmentation algorithm and a graph convolutional network-based automatic bullseye quantitation model. This novel system could provide more accurate and intelligent left ventricle auto-reorientation, localization, segmentation as well as more comprehensive and accurate automated quantification of the bullseye analysis. It would serve the clinical application of nuclear medical imaging and promote the development of nuclear cardiac diagnosis in China