CASIA OpenIR  > 毕业生  > 博士学位论文
肿瘤PET/CT成像影像组学相关算法的研究
牟玮
学位类型工学博士
导师戴汝为 ; 田捷
2016-05
学位授予单位中国科学院大学
学位授予地点北京
关键词正电子发射成像/计算机断层成像(Pet/ct) 肿瘤分割 肿瘤分期 生存 分析
其他摘要

      准确的判断肿瘤分期及评估预后对制定个性化的治疗方案具有重要的指导意义。 但目前常用的活检及外科检查有较高的主观性、有损性及错误率,因此研究一种定 量客观的诊断标准,搭建一套用计算机来辅助诊断的方法框架是非常有必要的。正 电子发射断层成像(Positron Emission Tomography, PET)作为一种分子影像学技术, 利用由放射性核素标记的示踪剂进行成像,能够在分子水平和细胞水平上反映肿瘤 的代谢功能信息,提供肿瘤的生理学信息,在肿瘤的早期诊断、临床分期、及预后 评估方面有非常重要的临床价值。

       作为临床上最常用的示踪剂,18F-FDG(2--2-脱氧-D-葡萄糖)可准确反映体内器官或组织的葡萄糖代谢水平,但是如果不对数据做进一步的分析,仅仅依靠图 像本身是无法对肿瘤的生理学信息进行准确的评测的。而借助图像的影像组学分析, 我们可以对图像进行处理和特征提取,得到有意义的量化指标,从而达到利用影像 学自动实现肿瘤的辅助分期及预后评估的目的。本文围绕着正电子发射成像/计算机 断层成像(Positron Emission Tomography/Computed Tomography,PET/CT)图像影像组学 (Radiomics)分析中存在的问题,以临床常发宫颈癌及肺癌为切入点,进行以 下几方面的研究。

(1)针对 PET 图像中宫颈癌肿瘤与膀胱难以区分这一困难,研究宫颈癌肿瘤的自 动分割算法。首先基于肿瘤和膀胱的组织特异性,利用 PET 图像和对应的 CT 图像构建超图,通过模糊聚类的方法确定初始肿瘤区域。然后利用 PET 图像中 肿瘤边缘和膀胱边缘的梯度场方向相反这一特点,将梯度场的信息集成到水平 集的框架中,构造新的演化方程,实现肿瘤的准确分割。该算法在仿真数据集 (7 种模型 5 种噪声水平)和 27 PET/CT 临床数据集上测试有效。

(2)针对目前常用的 SUVmax 具有一定局限性的缺点,寻找并提取非冗余、高重 现且具有丰富信息量的适用于临床定量诊断的新的特征参数,并根据提取的特 征参数寻找合适的模型实现宫颈癌患者在分期意义上的分类,即通过 PET 图像 实现宫颈癌患者临床分期的判断,从而辅助临床医生进行诊断并提供个性化的 治疗方案。该方法在 42 PET/CT 临床数据集上测试有效,并在仿真数据集(5 种重建参数 7 种噪声水平)下验证了参数的稳定性。 

(3) 针对单模态医学图像不能同时反映结构信息和功能信息,而现有影像组学分 析又多以单模态医学图像为主这一不足,研究融合图像影像组学分析在肺癌预 后评估中的意义。首先提出了一种基于改进的 Pansharp 模型的融合算法,更好 的保留了图像的细节信息。然后根据基于梯度场信息的改进的水平集算法和基 于 CV 模型的水平集算法提出了一种针对肺癌肿瘤的半自动联合分割算法。最 后在提取 PET CT 图像中的特征参数的同时构造了一系列基于融合图像的特 征,寻找反映肿瘤预后的特征参数及合适的分类模型实现早期肺癌患者在三年 存活率意义上的分类,并通过Kaplan-Meier 曲线及log-rank检验进行生存分析。 该算法在 42 PET/CT 临床数据集上测试有效。 


;

Accurate tumor staging and prognosis evaluation play an important role in treatment planning. However, the current diagnosis system, which usually depends on aspiration biopsy, physical examination and multiple imaging techniques, is always subjective, invasive and discrepant. Therefore, it is desirable to develop an objective and quantitative criteria to aid cancer diagnosis. As a molecular imaging technique, Positron Emission Tomography (PET) could reflect the metabolic characteristics of the tumors at the molecular level and provide physiological information by displaying the concentration distribution of the radioactive tracer, and has important clinical significance in the early diagnosis, staging and prognosis evaluation.

As a glucose analog and the most commonly used radioactive tracer, 18F-FDG (2-Deoxy-2-[18]F-Fluoro-D-glucose) could reflect the levels of glucose metabolism of human organs and tissue. However, it is difficult to evaluate the physiological information of the tumors accurately without the further analysis of the images. With the quantitative analysis of the images and the further mathematical process, we could obtain the meaningful quantitative indices to aid the staging and prognosis evaluation of the cancer victims. Therefore, this study focused on the radiomics analysis of Positron Emission Tomography/Computed Tomography (PET/CT) images, and carried out the work in the following aspects in terms of cervix cancer and lung cancer.

(1) Aiming at the difficulty of distinguishing cervix tumor and bladder in PET images, we proposed an automatic tumor segmentation algorithm. Firstly, a hyper-image was constructed with the PET images and the corresponding CT images, and the initial tumor region could be obtained with fuzzy clustering method based on the proposed tissue specificity. Then given the fact that the gradient fields of the boundary of the bladder and tumor should be opposite, we incorporated the gradient field information into the level set framework and constructed a new evolution equation to obtain the tumor automatically with high accuracy. The proposed method has been validated on simulated datasets (7 phantoms and 5 noise levels) and 26 clinical PET/CT images.
(2)  Aiming at the limitations of the commonly used SUVmax feature, we tried to extract and find non-redundant, highly reproducible and informative parameters applicable to quantitative clinical diagnosis, and find an appropriate classification modal to stage the cervix tumor automatically according to the features. Accordingly, we attempted to assist the precise diagnosis, and guide the personalized medicine in clinical. The proposed method has been tested on 42 clinical PET/CT images, and the robustness of the features has also been validated on simulated datasets (5 reconstruction parameters and 7 noise levels).
(3)  Aiming at the deficiency of the single modality medical images, we investigated the significance of multi-modality medical images in the prognosis evaluation of lung cancer patients. Firstly, we proposed a more efficient fusion method based on the improved Pansharp model to retain the detailed information of the original images. Then a semi-automatic joint segmentation method was proposed to segment the lung tumors both from PET images and CT images with the level set methods based on the gradient field and Chan-Vese (CV) modal. Finally, we extracted textural features both from PET images and CT images, and constructed a series of features based on the fusion images to find the appropriate parameters to reflect the recurrence and the prognosis through the analysis of survival curve using Kaplan-Meier method and log-rank test. The proposed method has been verified on 42 clinical PET/CT images. 


文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/11644
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
牟玮. 肿瘤PET/CT成像影像组学相关算法的研究[D]. 北京. 中国科学院大学,2016.
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
肿瘤PETCT成像影像组学析相关算法的研(8408KB)学位论文 暂不开放CC BY-NC-SA请求全文
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
[牟玮]的文章
百度学术
百度学术中相似的文章
[牟玮]的文章
必应学术
必应学术中相似的文章
[牟玮]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。