信息工程与人工智能学院知识库

      肺癌是一种全球范围内重要的癌症类型，其发病率和死亡率一直居高不下。尤其是非小细胞肺癌（Non-small Cell Lung Cancer, NSCLC）患者占肺癌病例的80%-85%。对于非小细胞肺癌患者来说，淋巴结转移在治疗方案制定中起着至关重要的作用。近年来，医学影像分析领域中深度学习技术的应用呈上升趋势，可以用于辅助淋巴结转移的分期工作。PET/CT作为一种多模式成像系统，结合了代谢和形态学信息，被广泛应用于分期评估。然而，获取和标注医学影像数据的成本较高，而且并不是所有患者都能接受被认为是分期金标准的病理检测。因此，影像诊断成为另一种选择，但其准确性受到多重因素的影响。目前已有一些研究提出了解决噪声标签问题的方法，但在医学领域中，对于联合噪声标签训练问题的研究较少。因此，本研究旨在运用半监督深度学习技术来帮助诊断患有非小细胞肺癌的患者是否存在淋巴结转移，为淋巴结转移的分期评估提供更有效的辅助工具。主要研究内容如下：

      （1）本文首先介绍了一种改进自 Mean Teachers 方法的师生模型，名为双学生师生模型（Dual Student with Teacher and Correction, DSTC），该模型包含两个学生网络和纠正网络，用于淋巴结转移预测。经过预处理的PET/CT肿瘤块作为输入，学生网络以病理和影像标签为目标，提取图像的深层次特征并预测淋巴结转移。在两个学生网络之间引入单向学习方法，以病理标签的指导下辅助影像标签预测淋巴结转移。教师网络的权重采用学生网络的指数移动平均值，通过一致性学习使得学生网络的预测与之相似。尽管放射科医生根据影像检查结果及相关疾病信息已经提供了诊断，为后续治疗和疗效评估提供了重要依据，但其准确度不如病理诊断。因此，在模型中加入了半监督策略纠正网络，以病理标签为指导，减少影像标签潜在的错误，纠正后的标签将用于学生网络的训练。该模型在具有328例影像标签的非小细胞肺癌患者的 PET/CT 数据集上进行了五折交叉验证，其中只有112例拥有病理标签。在测试集上，该模型实现了准确率为0.88和 AUC 为0.8，优于放射科医生的诊断和其他方法的预测，证实了本研究的模型能够有效预测非小细胞肺癌淋巴结转移。

      （2）由于医生手动勾画肿瘤区域的成本较高，本文还使用了一种无需肿瘤分割的多尺度图像淋巴结转移诊断模型。该模型以 PET/CT 图像的二维正面最大密度投影和三维半身图像作为输入，构建了一个多尺度、多模态的半监督深度学习模型，能够利用未带有病理标签的数据来生成伪标签并参与训练过程。该模型采用了两个基于 ResNet-18 的特征提取网络，用于提取二维和三维深度特征，每个特征提取网络包含 PET、CT 和堆叠三个分支。在堆叠的分支中引入了融合模块，以融合不同模态和上下文信息。最终将两个维度图像的特征进行拼接，用于预测淋巴结转移。实验结果表明，该模型在淋巴结转移分类研究中有效减少了对医生手动标注病灶区域的依赖，实现了分类准确度为 0.86。

      淋巴结转移作为疾病诊断过程中的关键环节，其准确性直接关系到后续治疗方案的制定以及患者预后的评估。因此，本文在淋巴结转移诊断领域所展开的研究，有望为医生提供更准确、更可靠的诊断依据，有助于改善患者的治疗效果和生活质量。

    Lung cancer is a significant type of cancer on a global scale, with high incidence and mortality rates. In particular, Non-small Cell Lung Cancer (NSCLC) patients account for 80%-85% of lung cancer cases. For NSCLC patients, lymph node metastasis plays a crucial role in treatment planning. In recent years, the application of deep learning techniques in the field of medical image analysis has been on the rise, providing assistance in staging lymph node metastasis. PET/CT, as a multimodal imaging system combining metabolic and morphological information, is widely used for staging evaluation. However, the cost of obtaining and annotating medical imaging data is high, and not all patients can undergo pathological testing, considered the gold standard for staging. Therefore, imaging diagnosis becomes an alternative, but its accuracy is influenced by multiple factors. While some studies have proposed methods to address noisy label issues, research on joint noisy label training problems in the medical field is limited. Therefore, this study aims to use semi-supervised deep learning approach to help diagnose whether patients with NSCLC have lymph node metastasis, and to provide a more effective auxiliary tool for the staging assessment of lymph node metastasis. The main research contents are as follows:

    (1) This paper first proposes a Dual Student with Teacher and Correction (DSTC) model, a modification of the Mean Teachers method, with two student networks and a correction network for predicting lymph node metastasis. Pre-processed PET/CT tumor patches serve as input, with the student networks targeting pathological and imaging labels to extract deep features from images and predict lymph node metastasis. A unidirectional learning method is introduced between the two student networks, to assist in predicting lymph node metastasis based on imaging labels under the guidance of pathological labels. The weights of the teacher network are updated using the exponential moving average of the student networks’ weights, ensuring similarity between the student networks through consistency learning. Despite radiologists providing diagnosis based on imaging results and relevant disease information, which serves as important evidence for subsequent treatment and efficacy assessment, its accuracy does not match that of pathological diagnosis. Therefore, a semi-supervised strategy correction network is added to the model, guided by pathological labels, to reduce potential errors in image labels. The corrected labels are then used for the training of the student network. The model is tested on a dataset of PET/CT images from 328 NSCLC patients, of which only 112 had pathological labels, through five-fold cross-validation. On the test set, the model achieves an accuracy of 0.88 and an AUC of 0.8, outperforming radiologists’ diagnosis and other prediction methods, confirming the model’s ability to effectively predict lymph node metastasis in NSCLC.

    (2) Due to the high cost of manual delineation of tumor regions by physicians, this paper also uses a multi-scale image lymph node metastasis diagnosis model without tumor segmentation. The model constructs a multi-scale, multi-modal semi-supervised deep learning model with two-dimensional frontal maximum density projections and three-dimensional half-body images of PET/CT images as inputs. It is able to utilize data without pathological labels to generate pseudo-labels and participate in the training process. The model employs two ResNet-18-based feature extraction networks for extracting two-dimensional and three-dimensional deep features. Each feature extraction network contains three branches of PET, CT, and stacking. A fusion module is introduced in the stacked branches to fuse different modal and contextual information. Finally, the features of the two dimensional images were spliced together to predict lymph node metastasis. The experimental results show that the model effectively reduces the dependence on doctors manually labeling the lesion area in the classification of lymph node metastasis, and achieves a classification accuracy of 0.86.

       Lymph node metastasis is a key link in the process of disease diagnosis, and its accuracy is directly related to the formulation of subsequent treatment plans and the evaluation of patient prognosis. Therefore, the research in the field of lymph node metastasis diagnosis is expected to provide doctors with more accurate and reliable diagnostic basis, which will help improve the treatment effect and quality of life of patients.