CASIA OpenIR  > 中国科学院分子影像重点实验室
基于影像组学的肝细胞癌预后因子预测方法研究
顾东升
Subtype博士
Thesis Advisor田捷
2021-05-17
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword影像组学,肝细胞癌,预后因子,无创诊断,深度学习
Abstract

    肝细胞癌(Hepatocellular Carcinoma,HCC)是全球范围内最常见的恶性肿瘤之一,其发病率和死亡率居高不下,严重危害人类健康。肝细胞癌预后因子的早期精准预测,对患者个体化治疗及随访方案的制定具有重要意义,有望降低患者复发率,提升患者生存。目前临床上对于肝细胞癌预后因子的精准诊断主要依赖于术前穿刺活检或术后的组织病理学检查。穿刺活检技术受限于肿瘤异质性,易出现采样偏差,无法对肿瘤精准评估,而组织病理学检查具有滞后性,难以为患者提供早期及时的治疗,并且二者都为有创的诊断方式,可能引起不良并发症。因此,亟需探索肝细胞癌术前无创精准诊断的新方法,辅助医生进行临床决策。

    计算机断层扫描成像(Computed Tomography,CT)和磁共振成像(Magnetic Resonance Imaging,MRI)作为临床常用的癌症分析工具,能够提供肿瘤及周围组织较为丰富的影像学信息,并且具有无创性,因此为癌症疾病的精准诊断带来了新的契机。然而,当前的影像学诊断主要依赖于医生的主观经验,具有不一致性,且无法挖掘肿瘤更深层次的信息。近年来,影像组学技术的快速发展为癌症疾病的术前精准诊疗提供了新的思路,在肿瘤的临床诊断和预后预测中已取得较好的进展。影像组学旨在从医学影像中挖掘大量反映肿瘤异质性的高维度特征,利用人工智能方法建立有效的预测模型,辅助临床医生进行癌症疾病的诊断和预后。本文以肝细胞癌Glypican-3(GPC3)分子标志物表达、微血管侵犯和病理分级等预后因子为临床诊断目标,研究CT和MRI影像组学技术在肝细胞癌术前诊断中的应用效果,并分别在图像输入阶段、特征筛选阶段及模型融合阶段对影像组学算法进行改进,提升模型对肝细胞癌预后因子预测的精度及泛化性能。本文的主要工作及贡献如下:

    1、提出了基于线性判别和信息熵的特征集成排序框架。本文针对肝细胞癌GPC3预后因子的无创诊断问题,探究在传统定量影像组学算法中,集成不同的特征排序策略对降低特征冗余、提升算法精度的有效性。首先根据Fisher 分数线性判别算法计算特征与标签的相关性,初步确定特征的重要性分数;然后采用最大信息系数相关性选择算法评估特征间的冗余;通过平滑非极大值抑制策略对二者进行集成并更新特征重要性排序,最后结合支持向量机分类器和序列前向选择算法得到最优的特征子集。最终,本研究提出的特征集成排序算法在GPC3表达预测中测试集受试者工作特征曲线下面积(Area Under Reciever Operating Characteristic Curve,AUC)可达0.915,相较于单阶段特征排序算法有较好的性能提升,同时相比其他的特征集成算法也体现出较好的优势。

    2、提出了基于双线性池化的多序列多模态深度迁移学习融合框架。本文针对肝细胞癌微血管侵犯预后因子的无创诊断问题,探究基于深度迁移学习的特征自动化分析方法及双线性池化融合策略对模型分类精度的提升作用。首先对单序列图像构建ResNet迁移学习网络,减少复杂的手工特征分析过程,同时提升模型的表达能力;然后对多序列MRI图像信息及临床病理信息进行双线性池化融合,通过建模特征的高阶统计信息来捕获特征之间的关系,进而生成更具表达力的全局信息。最终多序列多模态融合的迁移学习模型在测试集AUC为0.842,相较于定量的影像组学方法有显著的性能提升。此外,经过可视化的输入图像激活热图显示了与肝细胞癌微血管侵犯相关的病灶高危区域,使得深度学习模型在医学领域的应用更具可解释性。

    3、提出了结合双注意力感知的多尺度多通道密集连接网络模型(Multi-scale and Multi-channel Dense Convolutional Network,MSC-DenseNet)。本文针对肝细胞癌病理分级预后因子的无创诊断问题,探究在基于深度学习的影像组学建模中,多尺度多通道的肿瘤图像输入及注意力机制对模型判别能力的提升作用。首先构建多尺度的感兴趣区域输入图像,减少输入数据的组间方差;然后引入肿瘤、瘤周及混合区域三通道信息,充分利用病灶区域及病灶周围区域的异质性信息;最后引入空间位置和通道双重关系感知注意力模块来学习肿瘤图像不同位置微观表达的相关性和不同通道知识的相关性,自适应地整合局部特征和全局依赖。最终,在多中心数据集中MSC-DenseNet体现出最优的病理分级预测性能,AUC为0.813,相较于单尺度单通道图像输入有8%-10%的性能提升,对比实验表明该模型比临床和定量影像组学方法有更优的性能表现。

    相关成果以第一作者身份发表于Journal of Magnetic Resonance Imaging和 European Radiology期刊,以共同第一作者身份发表于Liver Cancer期刊。

Other Abstract

Hepatocellular carcinoma (HCC) is one of the most common malignant tumors in the world, and its morbidity and mortality rate remain high, which seriously endangers human health. Early and accurate prediction of hepatocellular carcinoma prognostic factors is of great significance for individualized treatment strategy making planning, which is expected to reduce the recurrence rate of patients and improve patient survival. Current clinical accurate diagnosis of hepatocellular carcinoma prognostic factors mainly depends on preoperative biopsy or postoperative histopathological examination. Needle biopsy is limited by tumor heterogeneity, sampling errors, which makes it impossible to accurately evaluate tumors. Histopathological examination are lagging and difficult to provide patients with early and timely treatment. Both biopsy and postoperative histopathological examination are invasive diagnostic methods, which would cause undesirable complications. Therefore, it is urgent to explore new methods for non-invasive and accurate diagnosis of HCC before surgery to assist doctors in clinical decision-making.

    Computed tomography (CT) and magnetic resonance imaging (Magnetic Resonance Imaging, MRI), as commonly used clinical cancer analysis tools, can provide rich imaging information of tumors and surrounding tissues. These radiological data are non-invasive acquired, so they bring new opportunities for HCC diagnosis. However, current imaging diagnosis mainly relies on the subjective experience of the doctor, which is inconsistent and cannot dig deeper information on the tumor, this brings great challenges to the accurate diagnosis of the disease. In recent years, the rapid development of radiomics has provided new ideas for the precise diagnosis and treatment of cancer diseases, and has achieved satisfying progress in the clinical diagnosis and prognosis of tumors. Radiomics aims to mine a large number of high-dimensional features reflecting tumor heterogeneity from medical images, and establish an effective prediction model based on artificial intelligence methods to assist clinicians in cacner management. The main work and contributions of this dissertation are as follows:

    1. A feature integration ranking framework is proposed based on linear discriminant and information entropy. This algorithm was designed to predict the expression of Glypican-3(GPC3) of HCC. In the process, we explored the effectiveness of feature integration ranking framework in reducing feature redundancy and improving prediction accuracy in traditional quantitative radiomics. First, the correlation between feature and label was calculated according to the Fisher score, and was initially defined as the importance score of the feature. Then, the correlation selection algorithm based on Maximal Information Coefficient (MIC) was used to evaluate the redundancy between the features; Next, the soft non-maximum suppression strategy integrated these two methods and updated the importance score ranking. Finally, the support vector machine classifier was ultilized to gradually iteratively obtain the optimal feature subset. After the feature integration ranking framework, the receiver operating characteristic curve of the test set in the discrimination of GPC3 positive and negative can reach 0.915, which was better than the single-stage feature ranking algorithm. It also shown better performance compared to other feature integration algorithms.

    2. A multi-sequence and multi-modal deep transfer learning fusion framework based on bilinear pooling is proposed in this dissertation. This algorithm was designed to predict the status of microvascular invasion of HCC. In the process, we explored the effectiveness of automatic feature analysis method based on deep transfer learning and the bilinear pooling fusion strategy in improving the classification accuracy of radiomics model. First, a ResNet transfer learning network for single-sequence image was constructed, which could alleviate the complicated manual feature analysis and improve the expressiveness of the model. Then the bilinear pooling was performed on the feature vectors of multi-sequence MRI images and clinical feature vector, this fusion strategy could capture the relationship and extract the high-order statistical information between different kinds of features, and then generate more expressive global information. Finally, the transfer learning model of multi-sequence and multi-modal achieved an AUC of 0.842 in test set, which had a significant improvement compared to the quantitative radiomics. In addition, the visualized activation heatmap for the input image shown high-risk areas of the lesion related to the microvascular invasion of HCC, making the deep learning model more interpretable in the medical field.

 3. A multi-scale and multi-channel densely connected network model (MSC-DenseNet) combined with dual attention perception was proposed in this dissertation. This algorithm was designed to predict the histopathological grading of HCC. In the process, we explored that if multi-scale and multi-channel tumor image and attention mechanism could improve the discrimination ability of deep learning model in radiomics analysis.First, multi-scale input image for region of interest were generated to reduce the variance among all the input data, and integrate the representation under multiple resolution. Then, by introducing the three-channel information of the tumor, peritumoral, and the mixed area, the heterogeneity of the  area in and around the lesion could be fully discovered. Next, the dual-relationship perception attention module of position and channel is introduced to learn the correlation of microscopic expressions in different positions of tumor images and the correlation of knowledge in different channels, which could adaptively integrate local features and global dependency. In the end, MSC-DenseNet showed the best prediction performance for histopathological grading in the multi-center datasets, with an AUC of 0.813, which is increased by 8%-10% compared with single-scale single-channel input image. The results also shown that this model yielded better performance than clinical model and quantitative radiomics.

    Relevant findings were published in the Journal of Magnetic Resonance Imaging and European Radiology as the first author, and published in Liver Cancer as the co-first author.

Pages122
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/44744
Collection中国科学院分子影像重点实验室
Recommended Citation
GB/T 7714
顾东升. 基于影像组学的肝细胞癌预后因子预测方法研究[D]. 中国科学院自动化研究所. 中国科学院大学,2021.
Files in This Item:
File Name/Size DocType Version Access License
基于影像组学的肝细胞癌预后因子预测方法研(19880KB)学位论文 开放获取CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[顾东升]'s Articles
Baidu academic
Similar articles in Baidu academic
[顾东升]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[顾东升]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.