基于图结构的消化道超声内镜图像分类算法研究

Endoscopic ultrasonography (EUS) has been widely used for the purpose of establishing medical image. In current diagnostic methods, the endoscopic ultrasonography(EUS) has greater diagnostic value which can obtain the clear internal structure of digestive organs. Compared with traditional imaging modalities, diagnostic accuracy of EUS is significantly higher than computed tomography(CT) and nuclear magnetic resonance imaging(NMRI). However, the diagnosis based on EUS images is affected considerably by the doctors' experience and subjective factors. Therefore, the development of an objective and quantitative diagnostic method based on EUS imaging for the early differential diagnosis is crucial to overcome these diagnostic difficulties. Graph based data has been proven powerful for representing pairwise relationship among features or objects. In this project, we focus on developing graph based approaches to EUS image feature selection and classification. For an EUS image, we commence by capturing the feature points from the image. With these feature points on hand, we thus establish a graph that represents the main characteristics of the EUS image. For the graphs abstracted from EUS images, we aim to develop novel graph based pattern recognition and machine learning algorithms for graph classification and analysis. These methods include a) graph kernels and b) graph complexity measure methods. Unlike most existing methods, our methods significantly overcome the shortcoming of neglecting vertex labels, requiring burdensome computation and reflecting limited information in a low dimensional space that arise in existing methods. Moreover, by integrating the developed methods, we also aim to develop a visualized system for the objective of deeply analyzing EUS images. This visualized system can create a variety of objective, quantitative diagnostic indices and an appropriate method to describe and understand the EUS images, and finally increasing the accuracy of early diagnosis of digestive system disease through EUS.

超声内镜(EUS)是目前广泛使用的体内成像技术，其清晰度超过很多传统的体外成像技术。如何结合计算机分析技术，予以EUS图像准确的诊断是一个热点问题。针对基于图结构的数据可以有效表达事物间结构关系的特点，本课题将采用图结构来对EUS图像进行特征表达，即，通过对EUS图像的特征点提取和特征点间的关系的计算来构建一个可以有效表达一幅EUS图像主要特征的图结构。其次，针对EUS图像构建的图，本课题还将设计新颖的、基于图论的机器学习与模式识别算法来对超声内镜生成的图像进行分类与分析。针对目前大部分基于图论的方法不能够被快速地计算、不能够抓取节点标记信息以及不能反映高维特征信息等缺陷，本课题将着重开发新颖的图的核算法和图的复杂度算法来对EUS图像的图结构进行深层次的结构与特征分析。以这些研究方案为基础，设计可供医生使用的可视化EUS图像诊断分析软件，从而对医生的医疗诊断提供客观的依据并提高诊断准确率。

1. 项目背景及科学意义：目前我国消化系统早癌的诊断率很低,就胃癌来说,日本可以早癌诊断率可以达到70%,韩国50%,而我国不到10%,远远达不到人们的要求,虽然随着医疗技术水平的提高,以及先进设备的应用为早癌的早诊早治提供了可能,但单纯依靠设备来提高诊断率是不现实的,怎样把各种设备进行整合,提高效率,充分利用计算机技术的客观性和准确性,实现诊断效能的飞跃是我们面临的重要课题..2. 主要研究内容：本课题还将着重使用新颖的基于图论的方法构进行深入的分析与处理，并提高对EUS 图像进行分析与分类的精度。具体来讲，结合数学、计算机和医学等知识，本课题将进行如下研究: ① EUS 图像的特征提取与相应的图的构建 ② 设计一个新的可以计算出带权重的图的复杂度算法； ③ 设计一个新颖的可以表达权重图特征的核算法：对于一个带权重的图，使用Perron-Frobenius 算法将其转化为一个有向的图。.3.该项目的研究已经取得了一些进展，主要成果有： ①研究并提出了若干种针对EUS图像的还原技术，能够将由于设备、操作、器官蠕动等因素导致的模糊的EUS图像还原为清晰图像，这有不但便于医生诊断，也有利于后续一些机器处理操作； ②提出了一种基于optimal surface方法的EUS图像分层技术，能够自动将消化道EUS图像中消化道的各层组织分割出来；③ 提出了一种全自动的病变区域检测及标注技术，能够自动检测给定消化道EUS图像中是否存在病变（例如早期癌症）。如果存在的话，能够自动将病变区域标注出来；.4. 该项目立项期间（2015-2017）：本人以第一作者或通讯作者发表文章18篇，其中JCR二区期刊2篇，三区期刊6篇，四区期刊1篇，CCF A类会议论文2篇，EI检索会议论文7篇。