毕业生知识库

胃癌是全球范围内最常见的恶性肿瘤之一，近年来新增的胃癌病例中约70%发生在亚洲，近一半发生在中国。及早发现、诊断及治疗是提高胃癌患者生活质量和改善预后的关键。胃癌的肿瘤-淋巴结-远处转移（Tumor-Node-Metastasis，TNM）分期系统描述了原发肿瘤侵犯程度（T分期）、区域淋巴结转移状态（N分期）及远处转移状态（M分期）三个方面的病情进展情况。临床中，及早准确预测胃癌病理TNM分期是分析肿瘤进展、评估患者预后以及制定个体化治疗方案的重要依据。

胃癌TNM分期的金标准是通过在显微镜下观察手术治疗过程中获取的人体组织病理样本确定的，具有侵入性和滞后性。对于胃癌患者的治疗前诊断，中国临床肿瘤学会和美国国立综合癌症网络发布的肿瘤临床实践指南均推荐将计算机断层扫描（computed tomography，CT）、磁共振成像、胃镜等无创的医学成像技术作为主要检查手段。其中，CT扫描具有成像迅速、分辨率高、可获取丰富人体内部组织信息等优点，是目前胃癌诊断中最常用、最主要的成像模态。然而，现阶段基于CT图像的胃癌TNM分期诊断方法主要依靠对胃癌原发灶及潜在转移部位的成像强度和形态学特征的评估，存在依赖医生经验、缺乏全球范围内统一的确切诊断标准、准确性不足、可重复性不高、费时费力等问题。因此，研究预测胃癌TNM分期的新方法，为医生提供稳定、精确、可靠的定量预测结果作为参考指标，是现阶段胃癌临床诊疗的迫切需求。

基于医学影像大数据和人工智能技术的影像组学方法为实现上述目标提供了新的途径。影像组学从医学图像中高通量地提取人眼无法获取的病灶信息，将肿瘤成像表型转化为高维特征，并深入挖掘图像特征与关键肿瘤临床指标之间的关联，进而构建具有定量预测能力的影像组学标签和模型。在已被报道的研究中，影像组学在肿瘤学领域的组织学诊断、分型分期预测、疗效评估以及预后风险分层等临床应用任务中均被证明有一定的预测能力。但是，现阶段还少有影像组学模型被应用于实际临床，这主要由于模型在预测精度、稳定性、可解释性和适用性等方面依然需要进一步提高，并且还需通过人群更广、功效更高、控制更严格的验证性实验加以评估。特别是在胃癌TNM分期预测任务上，对影像组学的相关研究尚处于起步阶段，过往研究主要以在小规模的单中心数据集上分析少量特征的临床显著性为主。因此，针对胃癌腹膜转移（M分期）预测、淋巴结转移（N分期）预测、浆膜侵犯（T分期）预测三个临床任务，本文分别根据各任务的内在特点和现阶段的研究难点开发了CT影像组学新算法，以提高预测的准确性和稳定性，推进影像组学在胃癌临床诊疗中的应用。我们在多中心数据集上对各预测模型进行了测试。本文的主要工作及贡献如下：

1、针对常规的CT主观征象在诊断胃癌腹膜转移时灵敏度不足的问题，本文受肿瘤转移的“种子-土壤学说”启发，提出了基于CT多区域特征的胃癌隐匿性腹膜转移预测算法。我们假设腹膜潜在的转移区域也含有可反映胃癌进展的早期征象，并以模型在临床中的可操作性和可解释性为导向，设计了胃癌腹膜转移的多区域特征提取框架，提出了同时量化胃癌原发灶和腹膜影像学表征的影像组学标签构建算法，进而结合关键临床指标建立了胃癌隐匿性腹膜转移预测模型。相比于单一区域影像组学，该模型融合了多区域信息，能够更加充分地量化与腹膜转移相关的表型特点，具有更高的预测精度。特别是在多中心测试集上，该模型对原本被基于CT主观征象的临床诊断所漏诊的腹膜转移阳性患者的预测准确性达到了85%以上，表明其能够在很大程度上降低腹膜转移患者接受不必要手术治疗的风险。由于算法具有显著的临床应用价值，基于该研究工作发表的论文已被连续三年（2019年、2020年、2021年）写入《中国临床肿瘤学会CSCO胃癌诊疗指南》。

2、针对临床对胃癌淋巴结转移精确诊断的迫切需求，并结合胃周淋巴结由于数量众多、分布广泛、解剖关系复杂等因素导致难以对其直接评估的实际情况，本文提出了基于CT深层特征的胃癌淋巴结转移预测算法，提高影像组学量化肿瘤原发灶表型的能力，提高特征提取对目标任务的适应性。该算法通过迁移学习及多阶段训练的策略引导深度学习网络学习到图像和临床任务之间的深层次映射关系，同时有效降低过拟合风险。为提取深度学习特征，我们使用正则化方法和相关性分析筛选出高信息量的非冗余节点，进而对节点输出的特征图进行多重测量，得到与任务显著相关的高通量特征。此外，我们设计了多阶段的特征选择方案，以进一步提高模型在多中心数据集上的稳定性。最终，我们构建了可以对胃癌具体N分期进行预测的影像组学模型，并在国内中心和国际中心构成的多个独立数据集上进行了测试。实验结果表明，与人工预定义特征、医生判断、现有临床指标相比，我们提出的基于CT深层特征的预测模型对胃癌淋巴结转移具有更好的预测效果。英国皇家马斯登医院临床研究主任David Cunningham教授在临床肿瘤学顶级期刊《Annals of Oncology》上撰写述评论文评价我们的工作“比临床N分期等临床指标更准确，可用于判断需要切除的淋巴结范围”。

3、针对目前胃癌浆膜侵犯缺乏稳定的精确诊断方法这一关键临床问题，以及多中心研究中CT成像参数的变化对影像组学特征存在较大干扰这一关键技术问题，本文提出了基于CT协调特征的胃癌浆膜侵犯预测算法。该算法使用一种基于“标准化-恢复”策略的图像协调框架减弱由图像层厚、像素尺寸、图像噪声等成像参数变化引起的图像变异。在该框架中，我们利用插值算法实现对不同中心CT数据三维分辨率的标准化，并使用一种基于多重注意力机制的图像恢复网络强化CT图像上感兴趣区域的关键表征。与其他图像预处理算法相比，该框架具有直接以图像作为输入输出、无需为不同成像参数额外收集训练样本、仅需训练单个网络模型的特点。在图像协调框架基础上，本文构建了融合无监督深度学习特征和预定义特征的胃癌浆膜侵犯预测模型。实验结果表明，我们的图像协调框架能够在提高模型泛化性能的同时，有效避免关键图像表征的损失。协调特征对肿瘤影像学表型具有代表性、对胃癌浆膜侵犯预测任务具有显著性。在多中心测试实验中，我们提出的基于CT协调特征的预测模型相比对照方法对成像参数变化较大的数据的预测效能提高了10.5%以上。

Gastric cancer is one of the most common malignant tumors worldwide. In recent years, about 70% of the new gastric cancer cases occurred in Asia, of which nearly half in China. Early detection, diagnosis, and treatment are the keys to improving the life quality and prognosis of gastric cancer patients. The tumor-node-metastasis staging (TNM staging) of gastric cancer describes the progression of the primary tumor invasion (T stage), the regional lymph node metastasis (N stage), and the distant metastasis (M stage). In the clinic, early and accurate prediction of pathological TNM staging of gastric cancer is an important basis for analyzing tumor progression, evaluating patient prognosis, and making optimal individualized treatment decisions.

The gold standard for TNM staging of gastric cancer is determined by the observation of human histopathological samples obtained during surgical treatment under the microscope, which is invasive and lagging. For the pretreatment diagnosis of gastric cancer patients, the Chinese Society of Clinical Oncology (CSCO) and the National Comprehensive Cancer Network recommend non-invasive medical imaging technologies such as computed tomography (CT), magnetic resonance imaging, and gastroscope as the main examination means. Among them, CT scanning has the advantages of rapid imaging, high resolution, and can obtain rich information on human internal tissue, and is currently the most commonly used and most important imaging modality in the diagnosis of gastric cancer. However, at present, the CT-based TNM staging method of gastric cancer mainly depends on the evaluation of the imaging intensity and morphological characteristics of the primary lesion and potential metastasis sites of gastric cancer. There are problems such as relying on doctors' experience, lack of accurate diagnostic criteria that are uniform worldwide, insufficient accuracy, low repeatability, and time-consuming and labor-intensive. Therefore, it is an urgent need for clinical diagnosis and treatment of gastric cancer to develop new methods for predicting the TNM staging of gastric cancer, so as to provide doctors with stable, accurate, and reliable quantitative prediction results as reference indicators.

The radiomics based on medical imaging big data and artificial intelligence technology provides a new way to achieve the above goals. Radiomics can extract high-throughput lesion information that cannot be obtained by the human eye from medical images, transform tumor imaging phenotypes into high-dimensional features, deeply excavate the correlation between image features and key clinical tumor indicators, and then construct radiomic signatures and models with quantitative prediction ability. In previous studies, radiomics has been proved to have certain predictive value in clinical application tasks, such as histological diagnosis, classification and staging prediction, efficacy evaluation, and prognostic risk stratification in the field of oncology. However, at this stage, few radiomic models have been applied to clinical practice. This is mainly because the models still need to be further improved in terms of prediction accuracy, stability, interpretability, and applicability, and it also needs to be evaluated by experiments with a wider population, higher efficacy, and stricter control. Especially for TNM staging prediction of gastric cancer, the relevant research on radiomics is still in its infancy, and the previous studies mainly focused on analyzing the clinical significance of a few features on a small-scale single-center dataset.

Therefore, for the three clinical tasks of peritoneal metastasis (M stage), lymph node metastasis (N stage) and serosal invasion (T stage) prediction of gastric cancer, this thesis respectively developed new CT radiomic algorithms according to the inherent characteristics of each task and the current research difficulties, so as to improve the accuracy and stability of prediction and promote the application of radiomics in the clinical diagnosis and treatment of gastric cancer. We tested each prediction model on the multi-center datasets. The main work and contributions of this thesis are as follows:

1. Aiming at the problem of insufficient sensitivity of conventional CT subjective signs in diagnosing peritoneal metastasis of gastric cancer, inspired by the "seed-soil theory" of tumor metastasis, this thesis proposes an algorithm for predicting occult peritoneal metastasis of gastric cancer based on CT multi-regional radiomic features. Hypothesizing that the potential metastatic region of the peritoneum also contains early signs that can reflect the progression of gastric cancer, and guided by the clinical operability and interpretability of the model, we designed a multi-regional feature extraction framework for gastric cancer peritoneal metastasis, and proposed a radiomic signature construction algorithm to quantify the image representations of primary lesion and peritoneum simultaneously, and then established a prediction model of gastric cancer occult peritoneal metastasis. Compared with single-region radiomics, the proposed model integrated multi-regional information, which could more fully quantify the phenotype related to peritoneal metastasis and had higher prediction accuracy. Especially in the multi-center test, the model obtained an accuracy of over 85% for the patients with positive peritoneal metastasis but missed by the CT signs, indicating that it could greatly reduce the risk of unnecessary surgical treatment for patients with peritoneal metastasis. Due to the significant clinical application value of the algorithm, the published thesis based on this research has been included in CSCO guidelines for the diagnosis and treatment of gastric cancer for three consecutive years (2019-2021).

2. Given the urgent clinical demand for accurate diagnosis of lymph node metastasis of gastric cancer and the fact that perigastric lymph nodes are difficult to be directly evaluated due to their large number, wide distribution, and complex anatomical relationship, this thesis proposes a prediction algorithm of lymph node metastasis of gastric cancer based on CT deep radiomics features, to improve the quantification ability of radiomics for the phenotype of the primary lesion and the adaptability of feature extraction to the target task. Through the strategies of transfer learning and multi-stage training, the algorithm guided the deep learning network to learn the deep mapping relationship between images and clinical tasks while effectively reducing the risk of overfitting. In order to extract deep learning features, we used the regularization method and correlation analysis to screen non-redundant informative nodes and then conducted multiple measurements on the feature map output by the nodes to obtain high-throughput features significantly related to the task. Furthermore, we designed a multi-stage feature selection scheme to further improve the stability of the model on multi-center datasets. Finally, we constructed a radiomic model that can predict the specific N stage of gastric cancer and tested it on international multi-center datasets. Compared with the hand-crafted features, doctor diagnosis, and existing clinical indicators, the experimental results showed that the proposed prediction model based on CT deep radiomics features had a better prediction performance on lymph node metastasis of gastric cancer. Professor David Cunningham, director of clinical research at the Royal Marsden, wrote a commentary paper in the top journal of clinical oncology “Annals of Oncology”, commenting this work as "more accurately than clinical N stage and other clinical characteristics, and could be useful to determine the needed extent of lymphadenectomy".

3. Aiming at the key clinical issue of the lack of stable and accurate diagnostic methods for serous invasion of gastric cancer, and the key technical issue that the changes of CT imaging parameters have a great interference on the radiomics features in multi-center research, this thesis proposes a prediction algorithm for the serous invasion of gastric cancer based on CT harmonization radiomic features. The algorithm used a normalization-restoration strategy-based image harmonization (NRIH) framework to reduce the image variations caused by the changes of CT imaging parameters such as slice thickness, pixel size, image noise, and so on. In this framework, we used an interpolation algorithm to standardize the three-dimensional resolution of CT data from different centers and used a multi-attention mechanism-based image restoration network to enhance the key representation of regions of interest in CT images. Compared with other image preprocessing algorithms, NRIH framework had the characteristics of directly taking the image as the input and output, without collecting additional training samples with different imaging parameters, and only training a single network model. Based on NRIH framework, we constructed a gastric cancer serosal invasion prediction model by integrating unsupervised deep learning features and hand-crafted features. Experimental results showed that our NRIH framework could improve the generalization performance of the model and effectively avoid the loss of key image representation. The harmonization radiomic features were representative of tumor imaging phenotype and significant for the prediction of serosal invasion of gastric cancer. In the multi-center test, the prediction efficiency of our prediction model based on the CT harmonization radiomic feature is improved by more than 10.5% compared with the compared method.