信息工程与人工智能学院知识库

肺癌仍是发病率和致死率最高的癌症之一，非小细胞肺（non-small cell lung cancer ,NSCLC）是最常见的肺癌类型，约占所有肺癌病例的85%；随着癌症的发展，肿瘤会从原发部位转移到身体其他组织，即出现转移。淋巴结转移是最常见的转移途径之一，是决定肿瘤分期和预后的主要因素。准确评估淋巴结转移状况对患者的治疗、预后和生存状况具有重要意义。目前对非小细胞肺癌淋巴结转移进行诊断的主要方法是病理检查，这是一种有创检查，并且可能引起患者的并发症，增加患者的经济负担。部分患者由于转移位置等原因无法获得病理检测结果。为了解决这一问题，本文提出了使用PET/CT影像对非小细胞肺癌患者的淋巴结转移状态进行诊断，期望为医生提供参考。主要研究内容如下：

（1）本研究利用多阶段深度学习技术，从PET/CT影像中识别NSCLC患者是否发生了淋巴结转移。该模型旨在结合部分患者的病理诊断结果和所有患者的放射科医生诊断结果，以识别NSCLC患者的淋巴结转移状态。在模型的第一阶段，我们使用内部训练集的有病理检测结果的数据集对模型进行预训练，并基于损失拟合高斯混合模型。第二阶段，利用缺乏病理结果但有医生诊断结果的数据集输入网络计算的损失和高斯混合模型的后验概率，从内部训练集中缺乏病理结果但有医生诊断结果的数据集中筛选出可靠的数据集。随后，我们结合有病理检测结果的数据和筛选出的可靠数据集对模型进行进一步训练。在第三阶段，引入5年生存率数据，以辅助评估不可靠数据的标签。最终，模型输出对淋巴结转移状态的诊断结果。该模型准确率为0.875，将模型结果辅助医生准确率达到0.911。

（2）为了对淋巴结转移状态进行深入预测细分为具体的N分期，本研究构建了深度学习框架。框架的目标是将正面和侧面的最大密度投影（Maximal Intensity Projection，MIP）特征以及影像组学特征整合在一起，以实现预测发生转移的NSCLC患者的具体N分期。研究中，网络采用ResNet18，深度学习特征的整合过程通过两个阶段的注意力机制进行整合优化。模型的预测性能通过深度学习和影像组学的特征结合而得到提升。实验结果显示，模型性能在增加了影像组学特征后得到提升，使用正面侧面特征融合模块也对模型性能提升有效。

Lung cancer remains one of the cancers with the highest incidence and mortality, and non-small cell lung cancer (NSCLC) is the most common type of lung cancer, accounting for about 85% of all lung cancer cases. As cancer progresses, the tumor metastasizes from the original site to other tissues in the body, that is, metastasis occurs. Lymph node metastasis is one of the most common routes of metastasis, and it is a major factor in determining tumor stage and prognosis. Accurate evaluation of lymph node metastasis is of great significance to the treatment, prognosis and survival of patients. At present, the main method for the diagnosis of lymph node metastasis of NSCLC is pathological examination, which is an invasive examination and may cause complications and increase the financial burden of patients. Some patients could not obtain pathological test results due to the location of metastasis. To solve this problem, this dissertation proposes to use PET/CT imaging to diagnose lymph node metastasis status in patients with non-small cell lung cancer. It is expected to provide reference for doctors. The main research contents are as follows:

(1) In this study, multi-stage deep learning techniques were used to identify lymph node metastasis in NSCLC patients from PET/CT images. The model was designed to combine pathological findings from some patients with radiologist findings from all patients to identify lymph node metastatic status in NSCLC patients. In the first stage of the model, we train the model using the data set with pathological detection results of the internal training set, and fit the Gaussian mixture model based on loss. In the second stage, data sets lacking pathology results but with physician diagnoses are used to filter the data according to the loss calculated by the data input network and the posterior probability of the Gaussian mixture model. Reliable data sets were screened from internal training sets that lacked pathology results but had physician diagnoses. The model was then further trained by combining the data with the pathologic findings and the selected reliable data sets. In the third phase, 5-year survival data was introduced to assist in the evaluation of unreliable data labels. Finally, the model outputs the diagnosis of lymph node metastasis status. The accuracy of the model was 0.875, and the accuracy of the model results assisted doctors reached 0.911.

(2) In order to deeply predict the status of lymph node metastasis and subdivide it into specific N stages, this study constructed a deep learning framework. The goal of the framework is to integrate frontal and lateral Maximal Intensity Projection (MIP) features with imaging omics features to predict the specific N stage of NSCLC patients. In this study, the network adopts ResNet18, and the integration process of deep learning features is optimized through a two-stage attention mechanism. By combining deep learning with radiomic features, we enhanced the model's predictive performance. The experimental results show that the performance of the model is improved after the addition of image radiomic features, and the use of front and side feature fusion module also has a certain effect on the performance of the model.