毕业生知识库

肺癌是危害人类健康的主要癌症，具有高发病率，低生存率的特点。因此，肺癌的早期诊断对于提高病人生存率尤为重要。然而，在早期诊断中，肺结节良恶性诊断往往需要依靠侵入式诊断，如活检或手术等。相对于非侵入式诊断，侵入式诊断会给患者造成更多痛苦。

 

近年来，计算机辅助诊断技术得到了快速发展，这项技术采用非侵入方式，通过分析患者病灶图像，辅助医生临床决策。在肺结节良恶性分类方面，利用肺部影像对肺结节进行定量分析面临着肺结节分割困难和特征提取耗时这两大难点。在研究针对肺结节的计算机辅助诊断算法时，需要考虑算法能否有效克服上述难点。卷积神经网络，作为一种图像分析框架，具有很强的图像描述性能。卷积神经网络采用端到端的训练方式，以原始图像为输入，图像标签为输出，简化了传统学习算法的流程。同时，网络通过监督学习的方式，自动提取高区分度特征。因此，在肺结节的良恶性分类方面，基于卷积神经网络的算法可以减轻对图像分割结果的依赖，降低特征提取的复杂度。

 

本论文主要研究了基于卷积神经网络的肺结节良恶性分类算法，包含针对肺结节恶性可疑度分析的基于多尺度结节图像的卷积神经网络的研究和实现，基于多区域池化的卷积神经网络的研究和实现，以及结节恶性可疑度的知识迁移几个方面。本论文的主要贡献如下：

 

1. 针对传统肺结节良恶性诊断算法中面临的肺结节分割困难和特征提取繁琐这两大问题，本论文提出了一种基于多尺度结节图像块的卷积神经网络算法。该算法以结节为中心，提取不同尺度的结节图像块，通过卷积神经网络提取不同尺度下的高度区分性特征，对肺结节恶性可疑度分类具有较好效果；

 

2. 针对高分辨率下肺结节图像的多尺度特征提取耗时这一问题，本论文提出了一种基于特征空间多区域池化的卷积神经网络算法。该算法对卷积特征的不同区域采用不同次数的池化操作，从单一尺度的卷积神经网络中提取多尺度特征，在对肺结节恶性可疑度分类达到较高正确率的同时，降低了计算量；

 

3. 针对医学图像分析领域中带有诊断信息的数据量较少，而卷积神经网络对数据量要求较大这一问题，本论文提出了一种基于迁移学习和多示例学习的病人级别肺癌检测算法，实现了对较大数据集的知识迁移以及与确诊数据集的知识融合，对病人级别肺癌检测具有较高的正确率。

 

Lung cancer is an aggressive cancer that risks people's health and it has a low long-term survival rate. Therefore, early diagnosis of lung cancer is important for cancer treatment. However, diagnosis of nodule malignancy based on biopsy or surgery is invasive which causes patients great pain when compared to non-invasive methods.

 

In recent years, computer-aided diagnosis has developed rapidly as a non-invasive method for diagnosis. Through analyzing lesion images, the computer-aided diagnosis system yields its report for clinical decision support. In the scenario of lung nodule malignancy classification, the traditional computer-aided diagnosis system faces two major challenges: the development of robust nodule segmentation method and effective malignancy relevant feature extraction. These challenges should be addressed when developing new computer-aided diagnosis algorithms.  As a powerful image analysis tool, the convolutional neural network (CNN), has been widely used in the field of image recognition and classification. There are several advantages using CNNs for image related analysis: a) the network is an end-to-end machine learning architecture where the input is the raw image and the output is the image label; b) due to the supervised learning scheme, discriminative features can be automatically learned. Therefore, in the scenario of lung nodule malignancy classification, CNNs can solve the two major challenges mentioned above. 

 

In this dissertation, we propose CNN based methods to perform lung nodule malignancy classification. Specifically, we propose a multi-scale CNN and a multi-crop CNN for lung nodule malignancy suspiciousness classification. We also investigate nodule malignancy suspiciousness knowledge transfer and adaptation in patient-level lung cancer detection. Our contributions are as follows:

 

1. Traditional lung nodule diagnosis systems face the challenges from nodule segmentation and feature extraction. In this dissertation, we propose to use a multi-scale CNN for lung nodule malignancy suspiciousness classification. The input images are multi-scale lung nodule patches. Features under each scale are extracted by a shared CNN and then are concatenated as the multi-scale features. We demonstrate that the proposed method can achieve high classification accuracy.

 

2. Multi-scale feature extraction is time-consuming when the nodule patch is large. In this dissertation, we propose a multi-crop CNN which crops different regions from convolutional feature maps and then applies max-pooling different times. Multi-scale features are extracted from a single scale network which reduces the computation cost. We demonstrate that the proposed method can also achieve high classification accuracy.

 

3. The performance of CNNs depends on the size of the dataset while diagnosis-definite medical datasets are usually very small. In this dissertation, we propose an integrated framework based on transfer learning and multiple instance learning to detect patient-level lung cancer.  The transferable knowledge is learnt from a large dataset and is then integrated with the knowledge from the diagnosis-definite dataset. We demonstrate that the proposed method performances well on patient-level lung cancer detection.