CASIA OpenIR  > 中国科学院分子影像重点实验室
多序列 MRI影像组学辅助肿瘤无创诊断的研究
魏靖伟
2019-05-30
页数161
学位类型博士
中文摘要

肿瘤的术前精确诊断对于患者的后续个体化治疗方案制定具有重要的指导意义,误诊或漏诊往往会使患者错失最佳的治疗机会,严重影响患者术后生存。目前的诊断方法多依靠术前活检或外科切除手术获得肿瘤组织进行确诊。然而,由于肿瘤的异质性,术前肿瘤穿刺活检会出现采样偏倚,无法对肿瘤进行全面且准确的评估;而通过获取术后标本进行病理组织学检查这一过程,具有滞后性,无法为患者的术前个体化精准治疗方案制定提供有效依据。因此,亟需研发可进行肿瘤术前无创诊断的新方法,以辅助医生临床决策。

随着成像技术的快速发展,医学影像在肿瘤诊疗中的作用日益凸显。由于影像可提供全面的视角和丰富的信息,且具有无创性,因此其为肿瘤的术前无创诊断带来了新契机。既往的定性影像分析虽可一定程度辅助术前诊断,但由于其无法定量精确的刻画肿瘤生物学特性,且往往依赖于医生的主观判断,因而无法为临床诊断提供最有效的依据支持。而基于医学影像大数据和人工智能算法的新兴技术——影像组学(Radiomics),为解决上述问题提供了新的思路。影像组学旨在从医学影像大数据中提取高通量、多尺度、多层次的量化特征,并结合影像征象、临床、病理或基因信息,使用人工智能技术构建综合智能分析决策系统,以解决肿瘤表型分型、疗效评估及预后预测等关键临床问题。而本文将利用基于多序列磁共振图像(Magnetic Resonance Imaging, MRI)的影像组学技术,围绕肿瘤“关键分子标志物”(基因)和“病理组织检测结果”(病理)的术前无创预测两个重要临床问题,开展以下几方面研究。

1)针对星型胶质瘤关键分子标志物MGMT启动子甲基化状态无法通过术前方法进行有效诊断这一关键临床问题,本研究提出了一种基于多栖息地的影像组学框架以实现对其的术前准确预测。相较于传统影像组学方法,该方法不仅提取了肿瘤实质区域的影像组学特征,还提取了瘤周水肿区域的定量特征,并深入挖掘肿瘤实质区域与瘤周水肿区域特征之间的关联性,综合纳入肿瘤实质区域及其微环境包含的关键影像组学特征组,进而进行影像组学融合标签的构建,实现MGMT启动子甲基化的术前预测。此外,本研究还进一步探寻了基于多栖息地的影像组学标签的预后价值,采用影像组学标签值对接受替莫唑胺化疗患者进行危险分层,并通过Kaplan-Meier曲线及log-rank检验进行生存分析,以筛选适合接受替莫唑胺化疗的患者群体。本方法在74例训练集和31例验证集数据上,受试者工作特征曲线面积(Area Under Reciever Operating Characteristic Curve, AUC)分别可达0.9250.902log-rank检验p值为0.03,证实了该方法进行星型胶质瘤MGMT启动子甲基化术前预测的有效性及其预后价值。

2)针对脑部肿瘤病理组织类型——颅内血管外皮细胞瘤和脑膜瘤难以通过传统影像分析进行术前无创诊断这一临床问题,构建基于T1WIT2WI图像的多序列融合影像组学模型以实现对其准确术前诊断。本研究提取并筛选具有鉴别能力和高度鲁棒性定量影像组学特征,并探寻基于小样本数据的最优建模策略,搭建并对比了64种建模方法以建立最优多参数预测系统,实现颅内血管外皮细胞瘤和脑膜瘤的术前无创预测。同时,本研究还实现了界面化的软件工具开发,为医生提供临床实践中可便捷使用的诊断工具,推动影像组学的临床转化。该模型在204例训练集和88例验证集数据上,AUC分别可达0.9850.927,证实了基于T1WIT2WI的多序列融合影像组学模型可实现脑部肿瘤病理类型的高准确度预测。

3)针对肝癌血管侵犯状态仅能通过术后病理检查进行诊断,无法实现术前评估和预测这一关键临床问题,本研究设计了具有肝癌特异性的剥落征和语义特征,构建基于普美显MRI的影像组学模型进行肝癌微血管侵犯状态的术前无创预测。基于本研究设计的肝癌影像组学特征,使用套索回归方法进行特征筛选和模型构建,并探寻基于普美显MRI的影像组学标签与基于计算机断层成像(Computed Tomography, CT)的影像组学标签对肝癌微血管侵犯状态进行预测的效能差异。本研究结果显示,在146例训练集和62例验证集上,构建的普美显MRI影像组学标签的AUC分别可达0.9430.861,该结果证实了基于普美显MRI的肝癌影像组学标签较基于CT的影像组学标签可更加准确地预测肝癌微血管侵犯状态,为肝癌患者后续个体化治疗方案制定提供了有效依据。

英文摘要

Accurate and preoperative diagnosis of tumors is essential for personalized treatment decision making. If patients are misdiagnosed or missed diagnosed, optimal treatment opportunity would be likely lost, severely affecting postoperative survival of patients. However, current diagnosis methods are mostly based on preoperative biopsy or postoperative tissue pathological examination. Due to tumor heterogeneity, preoperative biopsy may lead to sampling bias, thus fail to accurately and comprehensively evaluate the severity of disease. While postoperative tissue pathological examination could not provide timely and appropriate basis for preoperative treatment decision making. Thus, it is urgent to develop new methods for tumor preoperative and noninvasive diagnosis, assisting surgeons in clinical setting.

With rapid development of imaging techniques, medical image plays an increasingly prominent role in tumor management. Because medical images could provide full-scale and comprehensive information, meanwhile due to its noninvasiveness, it brings new chance for tumor noninvasive and preoperative diagnosis. Previous qualitative radiological analysis could assist preoperative diagnosis to certain extend, however, because it cannot quantitatively and accurately depict tumor biological characteristics, meanwhile highly depends on doctors subjective determination, it fails to provide reliable and effective basis for the diagnosis. By contrast, a newly emerging technique – radiomics that developed based on medical imaging big data and artificial intelligence techniques has the potential to solve abovementioned problems. Radiomics extracts high throughout, high-dimension, and multiscale quantitative features from medical images. Combining radiomic features with clinical, pathological or genetic information, an intelligent decision support making system would be established using artificial intelligent algorithms. It is widely used in phenotype classification, treatment evaluation, and prognosis prediction in tumor management. Therefore, this study utilized multi-sequence magnetic resonance imaging (MRI) based radiomics to improve the accuracy for preoperative and noninvasive diagnosis of tumors in prominent molecular marker prediction and postsurgical pathological examination outcome in terms of brain tumors and liver cancer.

(1)  To preoperatively predict a significant prognostic biomarker – oxygen 6-methylguanine-DNA methyltransferase (MGMT) promoter methylation in astrocytomas and validate its value for evaluation of temozolomide (TMZ) chemotherapy response, we proposed a multi-habitat radiomics pipeline. Besides exploring the radiomic features from tumor region, we additionally extracted features from peritumoral region, and analyzed the correlation between features from tumor and peritumoral habitats. The prognostic value of the established multi-habitat radiomics signature on overall survival for TMZ chemotherapy was explored using Kaplan Meier estimation and log-rank test. The multi-habitat radiomics signature exhibited supreme power for predicting MGMT promoter methylation with areas under receiver operating characteristic curve (AUCs) of 0.925 in the training cohort and 0.902 in the validation cohort. The radiomics approach could meanwhile divided patients into high-risk and low-risk groups for overall survival after TMZ chemotherapy with p-value of 0.03. It proved that the multi-habitat radiomics signature could accurately predict MGMT promoter methylation in patients with astrocytomas, and achieve survival stratification for TMZ chemotherapy, thus providing a preoperative basis for individualized treatment planning in clinical setting.

 (2) To perform accurate diagnosis of intracranial hemangiopericytoma (HPC) from meningioma by noninvasive method, we extracted discriminative and robust quantitative radiomic features from T1-weighted and T2-weighted MRI, and explored optimal pattern recognition strategies for small sample-based modelling. A multi-parameter model was established by combining radiomic features with tradition radiological characteristics and predictive clinical parameters for preoperative and noninvasive prediction of HPC and meningioma. Meanwhile, we developed a useful tool with user friendly interface for doctors to download with .exe file, which could make radiomics more practicable in routine clinical use. The proposed HPC and meningioma diagnostic tool achieved AUCs of 0.985 and 0.927 in the training and validation cohorts, respectively. It proved that the imaging and clinical information-gathered radiomics tool could realize diagnosis of HPC from meningioma with high accuracy.

 (3) To develop and validate a radiomics nomogram for preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC), we designed innovative radiomic features including peel-off features and semantic features that were specific for HCC noninvasive diagnosis, and optimal feature selection algorithm was investigated regarding to MVI prediction. Furthermore, we explored the diagnostic power of computed tomography (CT) and MRI for MVI preoperative prediction. The radiomics analysis proved that gadoxetic acid-enhanced MRI performed with better predictive ability for MVI noninvasive discrimination with AUC of 0.943 in the training cohort and of 0.861 in the validation cohort, which proved gadoxetic acid-enhanced MRI based radiomics model incorporating clinicoradiological risk factors and radiomic features could achieve satisfactory accuracy for MVI prediction in HCC.

关键词影像组学 人工智能 肿瘤 无创诊断 磁共振成像
语种中文
七大方向——子方向分类医学影像处理与分析
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/23955
专题中国科学院分子影像重点实验室
推荐引用方式
GB/T 7714
魏靖伟. 多序列 MRI影像组学辅助肿瘤无创诊断的研究[D]. 北京市海淀区中关村东路95号. 中国科学院自动化研究所,2019.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
0_CASIA Thesis-魏靖伟-明(3432KB)学位论文 开放获取CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[魏靖伟]的文章
百度学术
百度学术中相似的文章
[魏靖伟]的文章
必应学术
必应学术中相似的文章
[魏靖伟]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。