信息工程与人工智能学院知识库

    肺癌作为对生命威胁最大的癌症之一，发病率和死亡率一直以来都居高不下。而非小细胞肺癌占全部肺癌患病人数的85%。一般来说一旦发生了肿瘤转移，患者的生存期和预后也会变得更差。通过计算机辅助诊断是否发生淋巴结转移，可以以非侵入的方法计算出淋巴结发生转移的几率，进一步地给放射科医师提供必要信息从而提高诊断准确性，帮助改善病人存活期和预后。正电子发射断层扫描-计算机断层扫描（PET-CT）图像被认为是评估、分期、诊断肺癌的首选成像方式。影像组学是指从医学图像中提取和分析高维可挖掘数据，是癌症风险评估的重要预后工具。因此，本文从PET-CT图像中获取影像组学的信息，并用深度学习方法对非小细胞肺癌患者是否发生淋巴结转移做出了辅助诊断。主要研究如下：

     （1）基于深度学习的非小细胞肺癌淋巴结转移辅助诊断。本研究提出了一种端到端的深度学习体系结构，融合传统影像组学特征、深度学习特征、临床特征以及医生诊断进行淋巴结转移辅助诊断。首先从患者的PET图像中提取传统的影像组学特征，再用曼惠特尼U检验将存在显著差异的特征筛除；其次，分别利用三维卷积神经网络和二维卷积神经网络提取三维原发肿瘤和二维正面半身最大强度投影的深度学习特征；最后将传统影像组学特征和深度学习特征进行融合，再加入临床特征和医生诊断结果，对淋巴结是否发生转移做出诊断。本文在121例非小细胞肺癌患者的PET-CT影像和临床数据上进行研究。在测试集上该模型达到了0.82的AUC和0.86的精确度，明显优于医生诊断（AUC:0.61，精确度:0.80），具有更好的性能，证实了融合多种特征能以非侵入的方式辅助诊断是否发生非小细胞肺癌淋巴结转移。

     （2）两阶段的非小细胞肺癌淋巴结转移辅助诊断。本研究提出了一种两阶段的深度学习架构，通过分割任务使网络关注感兴趣区域，自动学习分类特征，消除测试阶段模型对医生勾画的金标准的依赖。首先在第一阶段使用一个分割网络对正面半身最大强度投影图像进行肿瘤分割，大致定位肿瘤区域，作为第二阶段三维肿瘤特征提取时的测试集处理依据；其次在第二阶段中利用三维分割-分类模型和二维分割-分类模型对三维肿瘤和二维正面半身最大强度投影提取深度学习特征；最后将两种深度学习特征融合，进行淋巴结转移辅助诊断。在测试集上该模型达到了0.74的AUC和0.84的精确度，证实了在测试集中不使用医生勾画的肿瘤金标准也能达到不错的性能。

       As one of the most threatening cancers to life, lung cancer has always had high morbidity and mortality. Non-small cell lung cancer (NSCLC) accounts for 85% of all lung cancers. In general, once tumor metastasis has occurred, the survival and prognosis of patients will also become worse. Computer-aided diagnosis of whether lymph node metastasis (LNM) occurs, the probability of LNM can be calculated in a non-invasive way, further providing radiologists with the necessary information to improve diagnostic accuracy and help improve patient survival and prognosis. Positron Emission Tomography -Computed Tomography (PET-CT) imaging is considered the imaging method of choice for assessing, staging, and diagnosing lung cancer. Radiomics refers to the extraction and analysis of high-dimensional mineable data from medical images, which is an important prognostic tool for cancer risk assessment. Therefore, in this paper, the radiomics features are extracted from PET-CT images and deep learning methods are used to assist in the diagnosis of whether LNM occur in patients with NSCLC. The main studies are as follows:

       (1) Computer-aided diagnosis of LNM of NSCLC based on deep learning. In this study, an end-to-end deep learning architecture is proposed, which integrates traditional imaging radiomics features, deep learning features, clinical features, and physician diagnosis for LNM computer-aided diagnosis. Firstly, the traditional radiomics features of the patient are extracted from the PET images of the patient, and then the features with significant differences in the image scanner are screened out by the Mann Whitney U test; secondly, the 3D convolutional neural network (CNN) and the 2D CNN are used to extract the deep learning features of the 3D primary tumor and the maximum intensity projection (MIP) of the 2D frontal bust, respectively; finally, the traditional radiomics features and deep learning features are fused, and then the clinical features and doctor's diagnosis results are added to diagnose whether the lymph nodes have metastasized. This paper conducts PET-CT imaging and clinical data from 121 patients with NSCLC. On the test set, the model achieved an accuracy of 0.82 AUC and 0.86, which was significantly better than the doctor's diagnosis (AUC: 0.61, precision: 0.80), and confirmed that the fusion of multiple features can assist in the diagnosis of NSCLC lymph node metastasis in a non-invasive manner.

       (2) Two-stage computer-aided diagnosis of LNM in NSCLC. This study proposes a two-stage deep learning architecture that makes the network pay attention to the region of interest by splitting the task, automatically learning classification characteristics, and eliminating the dependence of the test phase model on the gold standard outlined by the doctor. Firstly, in the first stage, a segmentation network is used to segment the tumor of the frontal bust MIP image, roughly localizing the tumor area, which is used as the basis for the test set processing in the second stage of 3D tumor feature extraction; secondly, in the second stage, the 3D segmentation-classification model and the 2D segmentation-classification model are used to extract deep learning characteristics from the 3D tumor and the 2D frontal bust MIP ; finally, the two deep learning features are fused to assist in the diagnosis of LNM. The model achieved an AUC of 0.74 and an accuracy of 0.84 on the test set, confirming that good performance could be achieved without the gold standard of tumors outlined by the doctor in the test set.