CASIA OpenIR  > 毕业生  > 博士学位论文
面向影像组学的图像配准及恶性肿瘤分类算法研究
喻冬东
学位类型工学博士
导师戴汝为 研究员 ; 田捷 研究员
2017-05-30
学位授予单位中国科学院研究生院
学位授予地点北京
关键词影像组学 图像配准 病理诊断 Egfr基因突变预测
其他摘要
       恶性肿瘤的早期诊断及其治疗效果评估,对于辅助医生制定有效治疗方案,提高患者生存时间以及生存质量具有重要的临床意义。然而恶性肿瘤往往具有空间异质性,这种异质性特点导致其预后效果较差。计算机断层成像(Computed Tomography, CT)作为一种结构成像技术,可以提供恶性肿瘤的空间异质性信息。准确衡量恶性肿瘤CT影像的异质性能够为其早期诊断及治疗效果评估提供关键信息。
       近年来,新兴的影像组学技术为恶性肿瘤的病理分型、临床分期、治疗效果评估等临床问题,提供了一种非侵入式的解决方案。它从临床影像数据中提取反映空间异质性的高通量定量影像特征,通过建立影像特征与临床信息之间的分类模型,辅助临床医生进行临床决策。相比传统的侵入式手段(活检、手术),影像组学技术可以更全面的解析恶性肿瘤的异质性,并且可以避免穿刺活检时给患者带来侵入式创伤。本文围绕计算机断层成像影像组学分析中存在的问题,以发病率和死亡率最高的肺癌为切入点,进行了以下几个方面的研究。
(1)   针对基于影像组学的疗效评估,需要首先解决治疗前后两个序列之间的空间位置对齐这一问题,本论文提出了一种基于高斯混合模型点集的大角度差异快速匹配算法。该算法首先使用并行加速多维尺度不变特征提取算法快速提取图像的特征点集。然后使用并行优化高斯混合模型点集匹配算法实现特征点集的配准。实验结果表明该算法将特征点的提取速度提高了200倍以上,并且提高了大角度差异图像的配准精度。
(2)   针对目前恶性肿瘤病理诊断主要依赖的活检标本病理检验有创性这一缺点,本论文提出了利用影像组学手段通过非侵入式的CT图像实现恶性肿瘤病理的诊断。该算法对恶性肺结节提取并寻找了具有丰富信息量的定量纹理特征,并根据这些特征参数建立合适的分类模型实现肺癌患者鳞癌腺癌的病理类型预测。实验结果表明该分类模型在临床数据集上取得了很好的预测结果(准确率=81.5%,AUC=0.91),为临床病理诊断提供了一种无创途径。
(3)   针对目前肿瘤的基因突变诊断主要依赖的活检标本病理检验有创性这一缺点,本论文提出了利用影像组学手段通过非侵入式的CT图像实现对非小细胞肺癌进行EGFR基因突变预测。该算法对恶性肺结节提取并寻找了具有丰富信息量的纹理特征和深度学习特征,并根据这些特征建立肺癌EGFR基因突变预测模型。实验结果表明该预测模型在训练集上取得了很好的效果(准确率=77.27%,AUC=0.83),并且在独立验证集上也有很好的泛化性能(准确率=75.00%,AUC=0.76),比Moffitt癌症研究所Gillies教授等人2016年发表在Radiology上的预测模型AUC值提高了8%以上。
 
关键词:影像组学,图像配准,病理诊断,EGFR基因突变预测
;          Early diagnosis of malignant tumors and early prediction of tumor response can help doctors select an effective therapeutic strategy to lengthen the patients’ survival time and improve their survival quality. However, malignant tumors tend to have spatial heterogeneity which leads to poor prognosis. As a structural imaging technique, Computed Tomography (CT) can provide the spatial heterogeneity information. Accurate assessment of malignant tumors’ heterogeneity can provide important information for early diagnosis and treatment evaluation.
        In recent years, the rapid development of radiomics provides a non-invase way for the pathological classification, clinical stage diagnosis, treatment assement and other clinical issues. It extracts high-throughput quantitative image features from clinical image data, and helps clinicians make clinical decisions by establishing a classification model between image features and clinical labels. Compared with traditional invasive methods (biopsy, surgery), radiomics can fully decode the heterogeneity of malignant tumors, and reduce the pain of tumor diagnosis. Therefore, this study focused on the radiomics of Computed Tomography (CT) images, and carried out the work in the following aspects in terms of lung cancer which is one of the most common cancers.
(1)  Aimed at the issue that image registration is need for the spatial alignment between two sequences before and after patient treatment, we propose a fast rotation-free feature based image registration method. First, the interest point sets from the pre-treatment and post-treament images are extracted using our accelerated-NSIFT method. Then, a Parallel Optimization based on the Gaussian Mixture Model Registration is proposed in order to obtain the transformation which best matches the interesting point sets. The experimental results show that the proposed algorithm can improve the extraction speed of interesting points by more than 200 times, and ameliorate the registration accuracy in the case of large pose differences.
(2)  Aimed at the issue that the pathological examination of malignancy tumor currently mainly depends on the biopsy which is invasive, we present a noninvasive radiomics-based CT image analysis for the pathological diagnosis. We extract and find informative texture feature parameters to represent each nodule. Based on these parameters, we established an appropriate classification model to classify the adenocarcinoma and squamous cell carcinoma. The experimental results show that the classification model has achieved good predictive results (Accuracy = 81.5%, AUC = 0.91) in the clinical data set, which provides a non-invasive way for clinical pathological diagnosis.
(3)  Aimed at the issue that the gene mutation status examination of malignancy tumor currently mainly depends on the biopsy which is invasive, we present a noninvasive radiomics-based CT image analysis for the prediction of EGFR gene mutation status in non-small cell lung cancer. We extract and find informative texture feature and deep learning feature parameters to represent each nodule. Based on these parameters, we established an appropriate classification model to predict the EGFR gene mutation status. The experimental results show that the classification model has revealed encouraging results in the primary cohort (Accuracy = 77.27%, AUC = 0.83) and achieved generalized performance in the independent validation cohort (Accuracy = 75.00%, AUC = 0.76). Our predictive model is more than 8% higher than the predicted model which proposed by Gillies from the Moffitt Cancer Institute in Radiology on 2016.
 
Keywords: Radiomics, Image registration, Pathology diagnosis, EGFR gene mutation status
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14641
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
喻冬东. 面向影像组学的图像配准及恶性肿瘤分类算法研究[D]. 北京. 中国科学院研究生院,2017.
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
博士学位论文喻冬东_修订版.pdf(6348KB)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
[喻冬东]的文章
百度学术
百度学术中相似的文章
[喻冬东]的文章
必应学术
必应学术中相似的文章
[喻冬东]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。