基于多模态深度特征融合表征肝细胞癌生物侵袭性研究

The biological aggressiveness of hepatocellular carcinoma (HCC) is of great significance to prognosis. Clinically, the biological aggressiveness of HCC are generally assessed by a small tissue obtained from the tumor by biopsy, which might not be much reliable due to sampling error. In addition, the relationship between the tumor and its adjacent parenchyma cannot be fully assessed, and the characteristics of the biological heterogeneity of neoplasm are unable to be fully evaluated from biopsy specimen as well. Recent studies have shown that texture features of arterial phase in contrast-enhanced MR imaging and Apparent Diffusion Coefficiency (ADC) in diffusion weighted MR are closely related to HCC staging and vascular invasion. However, the texture and ADC are low-level feature and limited to clinical experience, which may be insufficient to fully characterize the biological heterogeneity of neoplasm. Deep feature based on data-driven learning have consistently shown to be superior to traditional manual designed features in medical image processing. However, how to extract deep feature from clinical HCCs and how to establish the relationship between deep feature and malignancy or MVI of HCC have not been reported in the literature. This study proposes a systematic method to automatically extract deep feature from Contrast-enhanced MR using convolution neural network (CNN) in order to characterize malignancy and MVI of HCC. Meanwhile, the performance of deep feature for characterization of malignancy and MVI will be objectively compared with that of low-level morphologic features. This study also investigates a multi-kernel feature fusion method with support vector machine that can fuse deep features derived from multimodilities of contrast-enhanced MR images in a kernel space, which may yield better performance of characterization than that in a single modal for malignancy or MVI of HCC.

肝细胞癌（HCC）生物侵袭性表征对临床预后具有重要意义。临床上HCC生物侵袭性基于活检的病理切片显微图像确定，但易受到抽样误差的影响，且不能充分代表肿瘤与其相邻实质之间的关系和肿瘤异质性。最近研究表明，磁共振动态增强成像动脉期HCC纹理特征，及弥散加权成像的表面弥散系数，均与HCC分期和血管侵犯密切有关。但纹理特征和表面弥散系数属于影像低阶特征，并限于临床经验和人工设计，其可能不足以充分表征复杂的肿瘤异质性。基于数据驱动学习的深度特征，已显示出优于医学图像处理中的传统人工设计特征。但如何提取临床HCC深度特征，并建立深度特征与HCC分期和血管侵犯之间的关联，国内外研究尚鲜见报道。本研究拟基于卷积神经网络提取HCC深度特征表征临床分期和血管侵犯，并客观比较深度特征与纹理特征表征HCC侵袭性的差异。另外单模态表征可能存在局限性，拟通过多核函数支持向量机融合多模态深度特征，获得更佳的表征效果。

肝细胞癌(HCC)侵袭性表征对治疗决策和预后具有重要意义。HCC的恶性程度是预测手术切除或肝移植后复发和生存的重要预后因素,而微血管侵犯（MVI）已被证明是复发和预后不良的关键指标。本课题主要基于深度学习网络表征肝细胞癌侵袭性方面开展一系列的研究工作。针对扩散加权成像进行不同配准方法的运动校正，量化表明了不同配准方法对表征肝细胞癌恶性度的影响。基于多模态深度特征融合技术开展扩散加权成像和对比增强成像表征肝细胞癌恶性度的研究。探讨了二维和三维卷积神经网络在扩散加权成像和对比增强成像中预测肝细胞癌微血管侵犯性能。基于深度网络中局部和非局部特征的融合实现肿瘤的侵袭性表征，而非局部特征就是一种典型的注意力机制模型，实现了深度特征的全局信息提取，获得了比常规方法更优异的表征性能。基于超分辨率重建和自注意力机制克服了影像扫描层厚较大而导致表征性能降低的问题，在对抗网络中同时实现了分辨率提升和肿瘤分类性能的提升。基于深度网络中时间和空间信息的融合，实现了肿瘤微血管信息表征的性能提升。提出了可判别特征学习和自适应融合深度网络方法表征肝细胞癌恶性度。此外，通过DWI的多b值影像深度特征融合、生成对抗网络和自适应迁移技术、增强影像多模态的三维深度特征融合、二维和三维卷积神经网络深度特征融合、跨模态的深度网络迁移及生成对抗网络的样本虚拟生成等方面对肝细胞癌的侵袭性进行了深入的技术研究。在四年的研究过程中，我们发现扩散加权成像相比动态增强成像，由于其不使用造影剂在表征HCC侵袭性方面具有显著临床优势，基于3个b值（0,100,600）的扩散加权成像预测MVI性能（AUC=0.79）低于动态增强成像（AUC=0.926）。通过技术和临床深度结合，上述HCC侵袭性的一系列研究方法得到了专业领域内权威国际会议和国际期刊的认可，同时为临床HCC的侵袭性表征提供了智能化的诊断识别方法。