融合颈部淋巴结血流、弹性与结构信息的多模态超声诊断与疗效评估

It is of great clinical significance to diagnose and evaluate therapeutic effects for lesions of cervical lymph nodes (CLNs). Ultrasonic imaging can visualize the structure, blood flow of large vessels, blood flow of microvessels, and elasticity of CLNs by using multimodal techniques including B-mode, ultrasonic Doppler, contrast-enhanced ultrasound, and sonoelastography, respectively. It is valuable to fuse multimodal information for diagnosis and therapeutic effect evaluation of CLNs. However, it is challenging for multimodal information fusion due to large heterogeneity among multiple modalities of ultrasound. In order to overcome the obstacles of the heterogeneity in information fusion, in this proposal, we will first investigate the techniques of homogenization for multimodal data to make the noise levels, displacements, and cardiac phases consistent among modalities. Then we will segment multimodal images to delineate CLN lesions and typical regions within CLNs by using deep learning methods. We plan to acquire and fuse features from multimodal images so as to explore reliable imaging biomarkers for diagnosis and therapeutic effect evaluation. We also plan to implement the model of computer-aided diagnosis and therapeutic effect evaluation for CLNs with multimodal ultrasound, by using privileged information and multiple kernel learning. Finally, we will validate this model in clinical practices. It will offer a new tool for non-invasive diagnosis of CLN lesions, and a new means for non-invasive monitoring the treatment of lesions. The study can also be extended to quantitative analysis of diagnosis and treatment in other tissues or organs such as liver and breast.

颈部淋巴结(CLN)病变的诊断与疗效评估具有重要临床意义。包括B型、超声多普勒、超声造影、超声弹性成像在内的多模态超声成像技术分别关注CLN结构、大血管血流、微血管血流与弹性信息。有效融合多模态信息，对CLN诊断与疗效评估大有裨益。但多模态超声间具有强异构性，成为信息融合的巨大障碍。本课题针对强异构性CLN超声多模态信息的融合问题，首先研究多模态数据的同构化处理技术，使其噪声水平、位移、时相等尽量一致，接着基于深度学习分割多模态图像得到CLN病灶及其内部典型区域，然后获取与融合多模态CLN特征，挖掘可靠的诊断与疗效评估影像学标志物，继而运用特权信息、多核学习等方法实现CLN的多模态超声辅助诊断与疗效评估模型，最终检验其临床应用价值。课题将为无创诊断CLN病变提供新工具，为无创监测病变的治疗过程提供新手段。研究成果亦可推广至其它组织或器官(如肝与乳腺)诊疗中的定量分析，具有可延伸价值。

颈部淋巴结(cervical lymph node, CLN)病变的准确诊断与疗效评估具有重要临床意义。包括B型、超声弹性成像、超声造影、超声多普勒在内的多模态超声成像技术分别关注CLN结构、弹性、微血管血流与大血管血流信息。有效融合多模态信息，对CLN诊断与疗效评估大有裨益。但多模态超声间具有强异构性，成为信息融合的巨大障碍。本项目主要针对强异构性CLN超声多模态信息的融合问题进行探索，首先研究了多模态数据的同构化处理技术，使数据间的噪声水平、位移等达到尽量一致，接着基于深度学习技术分割了多模态图像得到CLN病灶及淋巴门等内部典型区域，然后获取与融合了多模态CLN的高维定量特征，进行降维并挖掘了可靠的诊断与疗效评估影像学标志物，继而运用特权信息学习、多项式深度网络、逐点门控深度网络等方法实现了CLN的多模态超声辅助诊断与疗效评估模型，最终检验了模型的临床应用价值。实验结果表明，基于多模态影像组学方法的四个分类诊断任务(良性/淋巴瘤性、良性/转移性、淋巴瘤性/转移性、良性/恶性)均达到了较好的诊断性能，受试者操作特性曲线下面积为0.960、0.716、0.933与0.856。基于特权信息学习的B型超声CLN良恶性诊断模型由于在训练阶段额外引入了其它模态信息，其分类敏感性与特异性提升至88%与77%，相较单模态B型超声训练的分类模型，两项指标均提高了8%。均一度与对比度等纹理特征在放疗前后的CLN转移灶中具有显著性差异(p<0.001)，能有效评估CLN病变的疗效。本项目有望为无创诊断CLN病变提供新工具，为无创监测病变的治疗过程提供新手段。研究成果亦便于推广至其它组织或器官诊疗中的定量分析，具有较好的可延伸价值。