多模态医学成像引导精准放疗中的快速知识协同与跨域知识借鉴学习方法研究及验证

Multimodal medical imaging-guided precision radiotherapy aims to incorporate multiple aspects of information in different types of medical images in order to assist doctors in more correctly assessing patients’ tumor lesions and true conditions of surrounding tissues/organs. Benefiting from such comprehensive information, doctors are capable of designing reasonable treatment plans as well as carrying out precise treatment procedures. However, some operations in existing precision radiotherapy either heavily depend on doctors’ clinical experience or are majorly carried out based on single-modal medical images. For addressing such challenges, the studies and trials of fast learning methods with knowledge collaboration and cross-domain knowledge reference will be pursued in this project. Particularly facing to the phases of precise diagnosis and precise localization, the key approaches with regard to object segmentation in the region of interest (ROI) based on fast multimodal knowledge collaboration and cross-domain knowledge reference, tissue/organ/tumor type recognition via multimodal knowledge collaboration and cross-domain feature enhancement, pseudo CT generation-based registration between multimodal medical images as well as PET attenuation correction, and retrievals of historically related images for objective medical images, will be studied and developed. So do some supporting algorithms of these approaches, such as multi-kernel fusion and knowledge reference-based collaborative spectral clustering and fast multi-kernel fusion-based supervised/semi-supervised collaborative classification. Moreover, via these studied contents, the precision radiotherapy-oriented multimodal medical image cloud will be built. The significance of our efforts in this project is two-fold: One the one hand, all of these studies are used to meet the direct technical requirements during the procedure of multimodal imaging-guided precision radiotherapy; On the other hand, some of the studies in this project are also compatible with the requirements of scientific research and clinical applications in other cases of precision medicine.

多模态医学成像引导精准放疗旨在结合各模态成像信息使医生较准确地评估患者肿瘤病灶及周边组织/器官真实状况，从而制订合理放疗方案、实施精确放疗过程。然而现行精准放疗部分环节要么较依赖医生临床经验要么主要面向单一模态成像进行。针对此现状，本课题拟开展快速知识协同与跨域知识借鉴学习方法的相关研究与验证，着重面向精确诊断和精确定位环节，研发关注区内(ROI)基于多模态快速知识协同和跨域知识借鉴的目标区域划分、基于多模态知识协同和跨域特征增强的组织/器官/肿瘤类型识别、基于生成模拟CT的多模态成像间配准与PET衰减校正、目标成像的历史关联成像检索等关键方法以及作为支撑的多核融合与知识借鉴协同谱聚类、多核融合有监督/半监督快速协同分类等算法，并基于这些内容构建面向精准放疗的多模态医学成像云。本课题工作既服务于多模态成像引导精准放疗的直接技术需求，也为其它精准医学相关课题的研究或临床应用提供通用技术支撑。

多模态医学成像引导精准放疗旨在结合各模态成像信息使医生较准确地评估患者肿瘤病灶及周边组织/器官真实状况，从而制订合理放疗方案、实施精确放疗过程。然而现行精准放疗部分环节要么较依赖医生临床经验要么主要面向单一模态成像进行。针对此现状，本课题开展快速知识协同与跨域知识借鉴学习方法的相关研究与验证工作，着重面向精确诊断和精确定位环节，为此项目组首先提出了基于知识极大化利用的半监督分类、一种多视角极大熵聚类、一种基于跨域共享潜在空间的抗负迁移模糊聚类、一种调整聚类假设联合成对约束半监督分类、基于域与样例平衡的多源迁移学习、流形学习与成对约束联合正则化非负矩阵分解等等基础算法，然后基于这些基础算法研发了一种基于迁移模糊聚类和主动学习分类的模拟CT生成、结合先验知识和部分监督的腹部骨盆从UTE-mDixon磁共振到CT的转换、一种由模拟图像引导的间接的多模态图像配准与修复、一种基于迁移模糊聚类和神经网络的PET/MR衰减校正等等核心技术，并基于这些内容构建面向精准放疗的多模态医学成像云。本课题的相关研究一方面为多模态医学成像引导精准放疗过程环节提供直接技术支撑，同时也为诸如医学成像辅助疾病筛查、医疗大数据挖据与利用等其它精准医疗相关课题的研究或临床应用提供通用技术支持。此外，本课题的研究内容也为社会其它行业的多视角大尺度数据挖掘问题也能提供重要技术借鉴价值。