信息工程与人工智能学院知识库

         肺癌是世界各地导致癌症相关死亡的主要原因之一。肺癌中，非小细胞肺癌 （Non-small Cell Lung Cancer, NSCLC)占肺癌病例总数的 80%-85%。近年来，由于分子生物学的发展，在 NSCLC 的治疗方法中，靶向治疗引起了广泛关注，特别是靶向酪氨酸的表皮生长因子受体(EGFR)激酶抑制剂(TKIs)已广泛应用于NSCLC 的治疗。与传统的治疗方法，如放疗、化疗等相比，EGFR-TKI 不仅副作用较少，且已被证明能更显著地改善发生了 EGFR 基因突变的 NSCLC 患者的预后，这说明了 EGFR 基因突变检测的重要性。 

       目前，EGFR 基因突变状态的识别主要是基于肿瘤标本活检的基因检测。然 而，在临床实践中，肿瘤的异质性和活检所得肿瘤组织的不足是造成无法准确检 测 EGFR 基因突变类型的主要原因。此外，活检检测还增加了患者癌症转移的潜 在风险，且不适合晚期肺癌患者和身体素质较差的患者。因此，开发一种基于 PET/CT 图像和临床特征的非侵入性且易于使用的方法来预测 EGFR 基因突变状 态在临床上具有较高的研究价值和应用前景。 

        本文拟采用深度学习的方法，利用患者的 18F-氟脱氧葡萄糖(FDG)正电子发 射断层扫描(Positron Emission Computed Tomography)/计算机断层扫描(Computer Tomography)（18F-FDGPET/CT)无创的预测患者的 EGFR 基因突变的状态。本文所做的研究主要围绕以下两个部分展开： 

        第一，由于现阶段针对 NSCLC 患者的 EGFR 基因突变状态预测的研究大多 是基于 CT 的单模态图像数据，而研究表明多模态数据能够提升深度学习模型的 预测能力。因此本研究首先使用了深度学习的预测方法，以 ResNet 网络作为主 干，采用创新的方法融合了 PET/CT 图像的深度学习特征和患者的临床特征，同 时与影像组学特征进行融合，来帮助网络提升预测性能。随后，本文进行了与传 统方法的对比实验，即基于临床模型、影像组学模型和综合模型预测 NSCLC 患 者的 EGFR 基因突变情况的预测。其中临床模型使用患者的临床特征，影像组学 模型使用 PET/CT 图像的双模态的影像组学特征，综合模型同时使用临床和影像 组学特征，最后利用传统的 AdaBoost、LogistRegression、SVM 等机器学习分类方法对 NSCLC 患者的 EGFR 基因突变情况进行预测。其中，影像组学模型首先需要对 PET/CT 图像数据进行影像组学特征的提取，并利用 RandomForest 方法兰州财经大学硕士学位论文基于PET/CT影像的非小细胞肺癌EGFR基因突变的预测对提取到的特征进行特征选择，以实现特征降维，缓解模型过拟合，提升模型预测精度；最后，将筛选后的特征输入到分类器中，得到最终的模型预测结果。实验结果表明，深度学习模型的预测结果相比传统的分析方法有较大的提升。第二，常用的深度学习预测方法的弊端之一是需要大量的具有病灶区域标注的数据，但医生手动标注的过程需要消耗大量时间、精力，并且不同医生之间难以保持绝对的一致性。因此，有病灶区域标注的医学影像数据极难获取。基于此问题，本研究提出了一种无需医生标注病灶区域进行 NSCLC 患者 EGFR 基因突变预测的深度学习方法，利用轻量级网络（MobileNet），将最大密度投影和均值密度投影作为模型的输入，无需医生标注便可得到与传统影像组学模型和临床模型相媲美的实验结果。

    Lung Cancer is one of the leading causes of Cancer-related deaths worldwide. In China, it is likewise one of the malignancies with the highest incidence and mortality rates. Among lung cancers, non-small cell lung Cancer (NSCLC) accounts for 80%-85% of all lung Cancer cases, and lung adenocarcinoma is the most common histological subtype of lung cancer, accounting for more than 40% of lung Cancer incidence. 

    In recent years, targeted therapies have attracted widespread attention in NSCLC due to the development of molecular biology, especially epidermal growth factor receptor (EGFR) kinase inhibitors (TKIs) targeting tyrosine have been widely used in the treatment of NSCLC. Compared with conventional therapies, such as radiotherapy and chemotherapy, EGFR-TKI has fewer side effects and has been shown to improve the prognosis of patients with EGFR-mutated NSCLC. Meanwhile, in patients without EGFR mutations, however, EGFR-TKI is relatively and may have a worse prognosis than platinum-based chemotherapy. This illustrates the importance of EGFR mutation detection. 

    Currently, the identification of EGFR mutation status is mainly based on genetic testing of biopsies of tumor specimens. However, in clinical practice, tumor heterogeneity and insufficient tissue obtained from biopsies are obstacles to accurately detecting EGFR mutation types. In addition, biopsy testing increases the potential risk of Cancer metastasis and is unsuitable for patients with advanced lung Cancer and those in poor health. Analysis of circulating cell-free tumor DNA (ctDNA) is another method to assess EGFR mutation status. Unfortunately, studies have shown that ctDNA testing has a relatively high false negative rate and is very expensive. Therefore, it is essential to develop a non-invasive and easy-to-use method to predict EGFR mutation status based on image and clinical features is essential. In this paper, we propose to use 18F-fluorodeoxyglucose (FDG) positron emission computed tomography (Positron Emission Computed Tomography)/ computer tomography (18F-FDGPET/CT) to predict the status of EGFR gene mutation in patients in a non-invasive way by using a deep learning approach. 

    The research done in this paper is centered on the following two components: 

    Since most of the current studies on EGFR mutation status prediction for NSCLC patients are based on unimodal image data from CT, it has been suggested that multimodal data may enhance the predictive power of deep learning models. In addition, it has also been shown that deep learning models can obtain better experimental results than imaging histology models in many studies targeting medical images. Therefore, this study first used a deep learning prediction method with ResNet network as the backbone and used an innovative approach to fuse deep learning features of PET/CT images and clinical features of patients, as well as a fusion of imaging histology features to help the network improve prediction performance. Subsequently, this paper conducted a comparison experiment with traditional methods, i.e., predicting EGFR gene mutations in NSCLC patients based on clinical, radiomics, and integrated models. The clinical model uses clinical features, the radiomics model uses radiomics features of PET/CT images, the integrated model uses both clinical and radiomics features, and finally, the EGFR gene mutations in NSCLC patients are predicted using traditional machine learning classification methods such as AdaBoost, Logist Regression, and SVM. Among them, the radiomics method firstly requires the extraction of imaging radiomics features from PET/CT image data and the feature selection of the extracted features using the Random Forest method to achieve feature dimensionality reduction, alleviate model overfitting and improve model prediction accuracy; finally, the filtered features are input to the classifier to obtain the final model prediction results. First, however, due to the difficulty of traditional clinical and radiomics methods to extract the deep features of PET/CT image data, the experimental results of deep learning significantly improve the results obtained compared with traditional analysis methods.

    Second, one of the drawbacks of commonly used deep learning prediction methods is that they require a large amount of data with lesion region annotation. Still, the manual annotation process by physicians consumes a lot of time and effort, and it isn't easy to maintain absolute consistency among different physicians. Based on this problem, this study proposes a deep learning prediction method for EGFR mutations in NSCLC patients without physician labeling of lesion regions. Using a lightweight network, the maximum density projection and the mean density projection are used as inputs to the model to obtain experimental results comparable to traditional radiomics models and clinical models without physician labeling.