多模态MRI图像肺癌检测与分类方法研究

The finding of the lung cancer in its early stage is critical to the effective treatment for the patients. Regular detection technique without hurt to human body is the key to early detection. As the magnetic resonance imaging (MRI) is with the characteristics of no radiation and multi-modality (dynamic contrast-enhancement, diffusion-weighted imaging, T2-weighted), applying MRI images to the early detection and diagnosis for the lung cancer are gradually focused. The research related to detecting and classifying lung cancer based on MRI images is still in the beginning stage and there exist the following problems. They were focused more on analyzing single modal MRI image information individually and less on fusing multi-modal MRI image information. The imaging mechanisms were taken less consideration in the feature extraction for each modal, which resulted in great errors in quantitatively describing the lung mass character. Moreover, the existing methods were more on analyzing the lung mass as a whole and less on analyzing its interior tissue at a deep level. Against this background, the study on the early detection and classification of lung cancer will be conducted in this project, by making full use of the feature information from multi-modal MRI images. The research contents include: the image registration for multi-modal MRI images, the lung mass segmentation, the feature extraction of the lung mass, feature fusion for multi-modal MRI images, and the benign and malignant classification of the lung mass on the basis of the multi-modal MRI images. This project refers to the theories and methods of information science and medical science. It is with important academic value and social benefits to explore the key technologies for detecting and diagnosing the lung cancers.

肺癌的早期发现对患者的有效治疗至关重要，而早期发现需要有对人体无损的定期检测手段。磁共振成像(MRI)对人体没有辐射，且具有多模态信息，其在肺癌早期检测与诊断中的应用逐渐受到关注。目前基于MRI图像对肺癌检测和分类的相关研究在国内外尚属起步阶段，存在以下主要问题：单独分析各模态MRI图像信息，缺乏多个模态MRI图像信息的融合；各模态特征提取方法缺乏对其相应MRI成像机理的深入分析，对肺肿块特性的定量描述存在较大误差；侧重肺肿块的整体分析，缺乏肺肿块内部结构的深层次分析。在此背景下，本课题充分利用多模态MRI图像信息，探究基于多模态MRI图像的肺癌早期检测与分类方法。研究内容包括：多模态肺部MRI图像配准，肺肿块分割，肺肿块特征提取，多模态特征融合，最后实现肺肿块良恶性分类。本课题结合信息学科与医学学科的理论和方法，研究肺癌MRI图像检测与分类的关键技术，具有重要的学术价值和社会效益。

肺癌是一种常见的恶性肿瘤，在全球范围内其发病率、死亡率极高。肺癌的早期发现对患者的有效治疗至关重要。磁共振成像（MRI）对人体没有辐射，且具有多模态信息，其在肺癌早期检测与诊断中应用逐渐受到关注。本课题构建了广东医科大学第一附属医院提供的肺部MRI图像数据库，并在该数据库上完成了：.（1）单模态肺部MRI图像配准：提出了基于多尺度局部刚性匹配的特征点对检测算法，并利用检测到的特征点约束非刚性肺部DCE-MRI图像配准，减少了配准中的非真实形变。为降低因错误匹配特征点对配准带来的负面影响，提出了基于自适应特征点约束权重的组配准算法，提高了配准DCE-MRI图像序列的精度。采用低秩矩阵分解算法对肺部DWI图像进行运动矫正，得到了噪声小、肿块边缘清晰的ADC图，提高了基于ADC的肿块良恶性分类的准确性和特异性。.（2）多模态肺部MRI图像配准：提出了基于地图集的多模图像配准方法与基于结构补偿的肺部多模MRI图像配准方法，解决了因多模图像间可能出现的信息不一致从而导致图像间失配的问题。提出了基于模态转换的多模图像预配准方法，相较于基于信息熵的预配准方法，提高了配准的速度。.（3）肺肿块检测与分割：提出了一种基于Faster R-CNN的磁共振图像肺结节检测算法，并利用肺结节解剖结构特征去除假阳性肺结节区域。提出了基于生成对抗网络的网络框架，实现了对肺部T2W图像中肺肿块的检测与分割。提出了一种基于全卷积网络和超密连接CNN模型的多模态图像分割方法，提高了与组织连接肿块的的分割精度。.（4）肺部肿块特征提取：对体素内不相关运动（IVIM）扩散权重磁共振成像在孤立肺部病变的诊断方面的应用进行了研究。研究表明IVIM参数提供了孤立肺病灶的功能信息，有助于良恶性病灶的鉴别与诊断。.（5）多模态特征融合与分类：提出了基于多参数磁共振成像的肿瘤良恶性分类方法，结果表明所提方法相较于单一参数MR图像序列具有更高的分类性能，基于多参数MR诊断方法具有较大的潜力。