基于PET/CT影像的非小细胞肺癌组织学亚型分类研究

作者	张杉杉
姓名汉语拼音	zhangshanshan
学号	2019000010013
培养单位	兰州财经大学
电话	18732622003
电子邮件	1347992943@qq.com
入学年份	2019-09
学位类别	学术硕士
培养级别	硕士研究生
学科门类	管理学
一级学科名称	管理科学与工程
学科方向	无
学科代码	1201
授予学位	管理学硕士
第一导师姓名	韩金仓
第一导师姓名汉语拼音	hanjincang
第一导师单位	兰州财经大学
第一导师职称	教授
题名	基于PET/CT影像的非小细胞肺癌组织学亚型分类研究
英文题名	Histological subtype classification of non-small cell lung cancer based on PET/CT image
关键词	非小细胞肺癌分类影像组学深度学习机器学习组织学亚型
外文关键词	Non-small cell lung cancer ; Classification ; Radiomics ; Deep learning ; machine learning ; Histological subtypes
摘要	肺癌是一种恶性肿瘤，它起源于人体支气管黏膜上皮或肺泡上皮。当前，肺癌的发病率仍然居高不下，且是全球癌症死亡的主要原因之一。临床上通常根据癌细胞的类型不同将原发性肺癌分为小细胞肺癌和非小细胞肺癌，非小细胞肺癌占据了原发性肺癌的80%以上，是原发性肺癌的主要构成部分。组织学亚型是非小细胞肺癌患者治疗方案制定的重要因素，术前区分非小细胞肺癌患者的组织学亚型，对于非小细胞肺癌患者的个性化诊疗有着重要意义。为了在术前更加准确、无侵入的区分非小细胞肺癌患者的组织学亚型，本文研究在包含 151 例组织学亚型为鳞状细胞癌或腺癌的非小细胞肺癌患者数据集上，建立了用于术前无创区分非小细胞肺癌组织学亚型的融合分类模型和全自动分类模型。在非小细胞肺癌组织学亚型融合分类模型中，本研究结合了传统影像组学获取的手工特征、深度学习方法构建的卷积神经网络获取的深度学习特征、非小细胞肺癌患者临床信息，使用机器学习分类器作为最终分类来区分患者组织学亚型。在这一模型中，本研究通过一系列对比实验证实融合模型在非小细胞肺癌组织学亚型分类任务较单独的影像组学和深度学习方法有着更好的性能（AUC 值 0.834，准确率=0.755、精确率=0.752、召回率=0.736 和 F1 值=0.740）。在非小细胞肺癌组织学亚型全自动分类模型中，本研究克服了非小细胞肺癌组织学亚型分类研究中对于医生手动标注的患者病灶区域的依赖性，构建了三种全自动分类模型，通过实验比较选择基于患者 PET/CT 影像最大强度投影的 2D 深度学习方法作为最佳全自动分类模型，并在这一全自动分类模型中得到了 0.768 的分类准确度。本文研究表明，传统影像组学和深度学习都能区分非小细胞肺癌患者的组织学亚型，其中融合传统影像组学获取的手工特征、深度学习方法构建的卷积神经网络获取的深度学习特征、非小细胞肺癌患者临床信息，使用机器学习分类器作为最终分类手段构建的融合分类模型较其他两种方法有着更大的优势。其次，脱离对医生手动标注的患者病灶区域的依赖，直接从患者医学影像入手判断其组织学亚型的非小细胞肺癌组织学亚型全自动分类方法是可行且有效的。
英文摘要	Lung cancer is a malignant tumor. Currently, the incidence of lung cancer remains high and is one of the leading causes of cancer deaths worldwide. In clinical practice, non-small cell lung cancer accounts for more than 80% of primary lung cancers and is the major component of primary lung cancer. Histological subtype is an important factor in the development of treatment plans for patients with non-small cell lung cancer, and differentiating the histological subtype of non-small cell lung cancer patients before surgery is important for personalized treatment of non-small cell lung cancer patients. A combined model and fully automated classification model for automatic and noninvasive distinguishing the histological subtypes of non-small cell lung cancer patients was established by Radiomics and deep learning methods on a dataset containing 151 non-small cell lung cancer patients whose histological subtypes were pathologically confirmed as squamous cell carcinoma and adenocarcinoma, in order to distinguish the histological subtypes of non-small cell lung cancer patients more accurately and noninvasively before surgery. The combined model of non-small cell lung cancer histological subtype classification was constructed by combining handcrafted features obtained from traditional Radiomics, deep learning features obtained from convolutional neural networks constructed by deep learning methods, and clinical information of non-small cell lung cancer patients using a machine learning classifier as the final classification tools. Using a series of comparative experiments, this study confirmed that this combined model performed better for the non-small cell lung cancer histological subtype classification task (AUC= 0.834, accuracy= 0.755, precision= 0.752, recall= 0.736, and F1 value= 0.740) and outperformed the results obtained using either traditional imaging histology or convolutional neural network models alone. In the fully automated classification model for non-small cell lung cancer histological subtypes, this study overcame the dependence on physician manually labeled patient lesion areas in non-small cell lung cancer histological subtype classification studies, constructed three fully automated classification models, and selected the 2D deep learning method based on the maximum intensity projection of patient PET/CT images as the best fully automated classification model by experimental comparison, and in this fully automated The classification accuracy of 0.768 was obtained in this fully automated classification model. This article shows that both conventional Radiomics method and deep learning techniques can distinguish histological subtypes of non-small cell lung cancer patients, in which the combined model built by fusing manual features obtained from conventional Radiomics method, deep learning features obtained from convolutional neural networks constructed by deep learning methods, and clinical information of non-small cell lung cancer patients using machine learning classifiers as the final classification tools has more advantages than the other two methods. Furthermore, it is feasible and effective for non-small cell lung cancer histologic subtypes to be classified directly from the patient's medical images, without relying on the physician's manual labeling of the patient's lesion area.
学位类型	硕士
答辩日期	2022-05-29
学位授予地点	甘肃省兰州市
语种	中文
论文总页数	52
参考文献总数	43
馆藏号	0004266
保密级别	公开
中图分类号	C93/72
文献类型	学位论文
条目标识符	http://ir.lzufe.edu.cn/handle/39EH0E1M/32329
专题	信息工程与人工智能学院
推荐引用方式 GB/T 7714	张杉杉. 基于PET/CT影像的非小细胞肺癌组织学亚型分类研究[D]. 甘肃省兰州市. 兰州财经大学,2022.

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
2019000010013.pdf（1277KB）	学位论文		暂不开放	CC BY-NC-SA	请求全文