毕业生知识库

  胃癌是全球范围内最常见的消化系统恶性肿瘤之一，其发病率和死亡率高居所有癌症的前五位，持续威胁国民生命健康。在根治性切除手术后，胃癌仍具有较高的复发率，术后5年生存率较低。因此，准确预测胃癌预后对患者个体化的治疗方案选择和随访策略制定具有重要意义，且有望提高患者的生存质量。目前，美国癌症联合委员会和国际抗癌联盟共同建立的肿瘤-淋巴结-转移（Tumor-Node-Metastasis，TNM）分期系统是临床上胃癌预后评估的主要依据，但该分期系统的提出依赖于回顾性的研究数据，具有一定的诊疗滞后性，需要不断更新完善。同时，胃癌具有高度的瘤内异质性和微环境异质性，仅依靠TNM分期难以准确反映患者的实际预后情况。因此，亟需探索精准有效的胃癌预后预测新方法，辅助医生进行临床决策。

  计算机断层扫描成像（Computed Tomography，CT）是临床上最常用于胃癌的无创诊疗手段，对确定肿瘤的位置、大小和浸润深度有重要的指导作用。传统的影像学评估主要依赖于医生对肿瘤组织形态、密度改变等生物学行为的简单估量和定性分析，易受到观察者主观经验的干扰，且通常不具备与肿瘤的病理类型和基因突变等微观信息的直接关联。近年来，影像组学技术的快速发展为医学影像分析领域带来了重大变革，为肿瘤的术前诊断和预后评估提供了有力的技术支持。影像组学旨在自动化地解析医学影像中高通量、多维度、定量化的肿瘤异质性表征，并利用人工智能方法建立影像表征与临床问题之间的智能预测模型，为临床医生提供辅助的患者管理工具。本文以胃癌预后预测任务为目标，研究影像组学技术在胃癌Borrmann分型和胃癌生存期预测中的应用效果，分别在特征学习和模型融合阶段对CT影像表征学习算法进行设计和改进，提升了预后预测模型的性能及泛化能力。本文的主要工作及创新点如下：

  （1）提出了基于多层感知机（Multilayer Perceptron，MLP）网络的机器学习模型集成学习框架。针对胃癌Borrmann分型缺乏术前无创精准诊断方法的问题，本文探究了基于特征工程的影像组学方法中，非线性集成不同机器学习模型的有效性。首先，从胃癌CT影像上提取预定义的影像组学特征，通过组合若干种特征选择方法和机器学习分类器，分别构建影像组学子分类模型；然后，设计了具有多个隐藏层和短路连接的MLP网络结构，通过反向传播算法进行模型训练，对多个子分类模型的预测值进行聚合，自适应地学习集成框架中各个子分类模型的权重。在Borrmann I/II/III vs. IV和Borrmann II vs. III两个诊断任务中，该算法的性能均优于单一的影像组学子分类模型，测试集中的受试者工作特性曲线下面积（Area Under Receiver Operating Characteristic Curve，AUC）可达0.767和0.768，且取得了更好的灵敏性和特异性结果，有效提升了预测模型的性能。

  （2）提出了基于多期相CT影像的对抗域自适应框架。针对胃癌无病生存期的术前预测问题，本文探究了深度域自适应方法对多期相CT影像之间共享表征学习的有效性以及对生存分析模型性能的提升作用。首先，基于ResNet结构搭建生存分析深度学习模型，端到端地学习影像特征并拟合其与患者生存风险概率的非线性关系，并在各源域（动脉期、门静脉期）的CT影像上预训练；然后，设计了具有非权重共享特征提取模块的多源域自适应网络，通过对抗训练的方式优化Wasserstein生成对抗损失函数，进而稳定地学习各源域-目标域CT图像对的域不变性图像表征；最后，提出了基于Wasserstein距离的加权策略，使特征分布更靠近目标域的源域获得更高的权重，实现了更准确的生存风险评估。该算法在目标域测试集中的一致性指数（Concordance Index，C-index）为0.668，较直接迁移、单源域自适应和多源域自适应等方法提高了10%~20%。此外，纳入临床分期特征后，模型性能得到进一步的提高，并在推荐接受新辅助治疗的患者亚组中实现了显著的风险分层（P = 0.0001），为医生提供了有效的辅助决策信息。

  （3）提出了基于自监督CT表征恢复的多任务学习网络。针对胃癌无复发生存期的术前预测问题，本文探究了基于自监督学习的图像恢复任务设计对CT影像表征学习的提升作用以及多任务学习策略对提高生存分析模型性能的有效性。该算法以生存分析为主任务，同时设计了基于自监督学习的拼图重组任务和遮挡图像修复任务，在没有生存数据监督训练时，这两个辅助任务可以从CT影像本身分别学习肿瘤的空间结构和语义特征。该算法还包括一种多任务学习网络的训练策略，网络的输入由原始图像、拼图图像和遮挡图像构成，并通过一个超参数调节每种图像的数据比例，使多任务学习网络可采用权重共享的编码器，并实现多个任务的共同优化。在胃癌无局部复发生存期和无远处转移生存期预测中，该算法在交叉验证的测试集中C-index分别为0.797±0.044和0.703±0.032，性能优于临床模型构建方法和基于特征工程的影像组学方法，且相较单一的生存分析任务有效提升了预测性能。

Gastric cancer is one of the most common malignant tumors of digestive system in the world with top five ranked morbidity and mortality rates, and continues to pose threats to human health. Patients who have received radical resection surgery are still confronted with a high recurrence rate, and the 5-year survival rate is relatively low. Hence, accurate prognostic prediction of gastric cancer is of great guiding significance for individualized treatment decision-making and follow-up strategy planning, which is expected to improve patients’ quality of life. Currently, the Tumor-Node-Metastasis (TNM) staging system established by the American Joint Committee on Cancer (AJCC) and Union for International Cancer Control (UICC) is the main basis for prognosis assessment of gastric cancer. However, this system relies on the research results of retrospective data, thus has a certain lag in clinical diagnosis and treatment and needs to be constantly updated. Meanwhile, due to the intratumoral and microenvironmental heterogeneity of gastric cancer, patients’ actual outcomes are hard to be fully reflected by TNM staging. Therefore, there is an urgent need to investigate new methods for accurate and effective prognostic prediction of gastric cancer to assist doctors in clinical decision-making.

    Computed tomography (CT) is a widely applied non-invasive tool for gastric cancer in clinical practice owing to its ability in identifying tumor site, size, and infiltration depth. Conventional radiological interpretation mainly relies on radiologists’ simple measurement and qualitative analysis of biological behaviors such as morphological and density changes in tumor tissues, which is easily interfered by individual perspectives of observers, and usually does not have direct correlations with the microscopic information of tumors such as pathological types and gene mutation status. Nowadays, the rapid development of radiomics has brought great changes to medical image analysis, and provides strong technical support for preoperative diagnosis and prognostic evaluation of cancers. Radiomics is designed to automatically capture the high-throughput, multi-dimensional, and quantitative feature representation of tumor heterogeneity, and establish prediction models by artificial intelligence techniques to assist clinicians in cancer management. This dissertation focused on the efficacy of radiomics in prognostic prediction of gastric cancer. Improved CT image representation learning algorithms were proposed in feature learning and model fusion stages, respectively, yielding better prognostic prediction performance. The main work and contributions of this dissertation are as follows:

    (1) A multilayer perceptron (MLP) network based multi-model ensemble learning framework was proposed for specific Borrmann classification in gastric cancer. This dissertation explored the validity of nonlinear integration of different machine learning models in feature engineering based radiomics. First, pre-defined radiomic features were extracted from CT images, and basic classification models were respectively constructed by combining different feature selection methods and machine learning classifiers. Then, an MLP network was designed with multiple hidden layers and shortcut connections and trained by back propagation algorithm. The basic classification model predictions were aggregated and the corresponding weights were adaptively learned. In Borrmann I/II/III vs. IV and Borrmann II vs. III tasks, the MLP ensemble model showed superior performance to basic classification models, achieving area under receiver operating characteristic curves (AUCs) of 0.767 and 0.768, respectively, with better sensitivity and specificity results. This framework effectively improved the prognostic model performance.

    (2) A multi-phase CT based adversarial domain adaptation framework was proposed for disease-free survival prediction in gastric cancer. This dissertation explored the effectiveness of deep domain adaptation methods in learning transferable domain-invariant feature representation of multi-phase CT images and improving survival analysis model performance. First, a survival analysis deep learning network was built based on ResNet and pre-trained on source domains with arterial phase and portal venous phase CT images, respectively. This network learned image features end-to-end and fitted the survival risk model nonlinearly. Then, a multi-source domain adaption network was designed, the feature extraction module of which adopted unshared weights. In this stage, the network was trained in an adversarial manner by measuring Wasserstein distance to stably learn the domain-invariant features between each source-target image pair. Finally, a Wasserstein distance based weighting strategy was proposed to combine the predictions from different source domains. This algorithm achieved a concordance index (C-index) of 0.668 in test set, increased by 10%~20% compared with direct transfer method as well as some single-source and multi-source domain adaption methods. Besides, incorporating clinical staging information further improved the model performance. The combined model could significantly risk stratify patients recommended to receive neoadjuvant therapy (P = 0.0001), which could provide auxiliary information for clinicians in patient management.

    (3) A multi-task learning framework was proposed based on self-supervised CT image restoration for recurrence-free survival prediction in gastric cancer. This dissertation explored the promotion of self-supervised image restoration tasks in representation learning of CT images and the effectiveness of multi-task learning strategy in improving survival analysis model performance. Survival analysis was the main task, and two self-supervised learning based auxiliary tasks were designed to learn spatial positions and semantic features of tumors in the absence of supervision by survival data, including a jigsaw puzzle reassembly task and a blocked image inpainting task. The algorithm also provided a training strategy for multi-task learning network: the inputs consisted of original images, shuffled images, and blocked images, whose proportions were adjusted through a hyperparameter. This enabled the multi-task learning network to use encoders with shared weights and be optimized jointly. In prediction of locoregional recurrence-free survival and distant metastasis-free survival, the algorithm achieved C-indices of 0.797±0.044 and 0.703±0.032 in cross-validation, respectively, outperforming the clinical model and feature engineering based radiomic model. Also, the algorithm effectively improved the model performance compared with single survival analysis task.