CT能谱信息及其特征选择与模式识别在胰腺肿瘤诊断及预后评估中的方法与应用研究

Pancreatic neoplasms consist of a number of heterogeneous tumors with different biological behaviors. Different kinds of tumor have various clinical characteristics and prognosis; even tumors with the same phenotype may have different invasiveness, and should be treated respectively. Ki-67 protein plays an important role in the development of pancreatic neoplasms. Early detection and accurate staging is the key of the clinical and research work for pancreatic neoplasms, which provides guidance to prompt and appropriate therapy. Although MDCT is the preferred imaging modality for the diagnosis of pancreatic neoplasms and is widely used in the clinical setting, there are certain limits considering early detection, differentiation, and accurate staging. Spectral imaging is the recent innovation of CT technique; it is a promising development, which bears the potential to improve lesion detection and characterization in comparison to the conventional CT techniques. But the amount of data derived from CT spectral imaging can be overwhelming even for expert readers. In recent years, pattern recognition and machine learning methods have been proved useful for diagnostic decision-making problems in high dimensional feature space. This study attempts to build a differentiation diagnosis model based on CT spectral imaging features of pancreatic neoplasms, and to assess the value of CT spectral imaging features and differentiation model for the diagnosis of pancreatic neoplasms. By using feature selection, this study tries to clarify the Ki-67 related spectral CT features. The study will involve the early detection of pancreatic carcinoma and insulinoma, differentiation diagnosis of pancreatic neoplasms, and staging of pancreatic carcinoma. The purpose of this study is to improve the sensitivity and accuracy of diagnosis for pancreatic neoplasms and to help early and proper treatment and prognostic evaluation, and also to provide the technical paradigm for future studies.

如何早期发现肿瘤、术前做出准确而全面的诊断从而使胰腺肿瘤能够早期治疗、合理治疗是临床工作与研究的关键问题。增殖细胞核抗原Ki-67是胰腺肿瘤发生、发展过程中的一个重要因子。能谱成像是CT领域的最新技术革新，最近的研究表明CT能谱成像有望提高病灶检测和诊断的敏感度和准确度，但其所获得的高维数据信息给常规诊断模式和思路提出了前所未有的挑战。机器学习与模式识别技术是处理和分析高维数据信息的有效方法，广泛应用于医学信息学领域。本研究试图对CT能谱信息进行特征选择、并通过模式识别和机器学习构建针对胰腺肿瘤的鉴别诊断模型；通过特征选择明确与Ki-67相关的CT能谱特征。研究内容主要包括胰腺癌和胰岛素瘤的早期检测、胰腺肿瘤的鉴别诊断及胰腺癌的术前分期，以期达到提高胰腺肿瘤诊断的敏感度和准确度，实现早期诊断、准确诊断的目的，为早期、合理治疗及预后评估提供全面可靠的信息；也为今后其它病变的研究提供方法学基础

项目的背景：CT能谱成像作为新的临床成像技术、提供了大量定量化信息特征，如何采用有效的方法将这些定量数据用于临床、协助胰腺肿瘤的诊断、鉴别诊断及研究是该项目研究的初衷。.主要研究内容：CT能谱成像多参数定量分析结合机器学习模型用于胰腺肿瘤的检测、鉴别诊断及病理、预后相关分析。.重要结果：能谱CT最佳单能量成像能够提高胰腺肿瘤的图像质量，有助于胰腺癌病灶的显示；最佳单能量成像结合碘基图像能够提高胰岛素瘤的检出率、其检出率与MRI、ASVS相当。在胰腺黏液性囊腺瘤（MCNs）与寡囊型浆液性囊腺瘤（SOA）的鉴别诊断中，CT能谱成像定量特征信息、支持向量机（SVM）结合基于Fisher的特征选择方法可起辅助作用。CT能谱成像定量分析可用于鉴别低级别与非低级别胰腺神经内分泌肿瘤（PNENs）；多参数联合诊断的效能高于单参数。CT能谱成像定量参数与肿瘤增殖指数（Ki-67）之间具有相关性。.关键数据：最佳单能量成像能够提高胰腺癌（113例）病灶对比噪声比>10％。SVM＋基于Fisher的特征选择方法结合CT能谱成像定量参数鉴别MCNs（19例）与SOA（23例）的准确率达93.02%。CT能谱成像多参数Logistic回归模型区分非低级PNENs（20例 vs 64例）的准确率为82.3%。.科学意义：基于SVM的机器学习模型能用于临床胰腺肿瘤的CT能谱成像数据处理， 筛选出合适的指标、并提高鉴别诊断的效能。