影像组学智能分析理论与关键技术

With the development of precision medicine, as one of the most important parts, intelligent analysis of medical images is attached more and more importance. Under the background of image-acquisition multi-modality, radiomics analysis and model construction intellectuality, this project, observing the analysis nature of ‘medical serving’, focus on the research of radiomics analysis under the dual standards of ‘computational intelligence’ and ‘clinical effectiveness’. First of all, according to the clinical characteristics of the images, the project focuses on the mechanism of medical risk generation during the process of intelligent computing, and classifies the risks in both direct and indirect ways. Secondly, we are going to do a statistical ensemble validating extractive research on the contradiction between intelligent computing and clinical implications of high-throughput features. Thirdly, aiming at the homology of image registration and segmentation, the project conducts reinforcement-learning research between two kinds of tasks based on multi-modal image data. Fourthly, based on the fact of the small scale and the uneven distribution of of medical images samples, we introduce transferring learning method, which is an important part of machine learning to the study of the mechanism of transfer analysis based on disease and image modality. Finally, based on the medical image data of prostate tumor, research of constructing dynamic on-demand visual platform will be done in this project. The research of this project attempts to construct a doctor-brain-like model, and is a great innovation of artificial intelligence in medical area.

随着精准医疗的大力推进，作为其重要内容之一的医学影像智能分析也全方位备受重视。本项目在影像采集多模化、影像分析组学化、影像建模智能化的背景下，坚守分析的“服务医疗”本质，开展“计算智能”与“临床有效”双重标准下的影像组学分析研究；首先，针对影像的临床属性，研究智能计算中的医疗风险产生机制，对直接与间接方式下的风险进行分级建模；其次，针对高通量特征的智能计算与临床内涵间的矛盾，开展统计集成的验证式提取研究；第三，针对影像配准与分割的同源性，基于多模态影像数据，开展两种任务间的强化学习研究；第四，基于医学影像的小样本性与样本分布不均的事实，引入机器学习中的迁移学习机制，研究基于疾病与影像模态的迁移分析机制；最后，基于前列腺肿瘤影像数据，开展肿瘤的动态按需可视化平台构建研究。本项目的研究试图让影像组学智能分析过程模拟医生诊断，是类医脑建模的一次尝试，更是人工智能在医疗领域的深度创新。

随着人工智能及健康中国两大国家战略的深入开展，基于医学影像组学人工智能技术研发越来越受到学界与业界的重视。本项目通过对影像组学分析中的数据多模态、不平衡、小样本及黑盒子等难点，开展了一系列研究。首先，针对数据的多模态，提出了一系列新的特征融合与强化学习算法，如基于结构张量结合特征驱动的多模图像配准、融合语义的卷积循环神经网络等，有效提升了模型的计算准确性与分析稳定性。其次，针对数据的不平衡及小样本问题，开展了迁移学习医疗影像分析研究，包括基于域适应的迁移学习算法、基于局部特征保持的模型等，有效提高了不平衡及小样本下的分析效果。最后，开展了高通量特征提取和医疗风险评估研究，提出了一系列高通量特征提取算法，包括基于稀疏公共特征表示的不平衡样本特征提取算法、跨模态场景下的图像检索信息表征算法以及一系列动态异质大图数据社区挖掘的基础理论与算法，有效提升了医疗影像特征在下游任务中的应用效果。上述理论创新在前列腺医学影像数据集上进行实现，形成了一套前列腺可视化分析系统，并在北医三院开展临床测试。本项目的研究，一方面提高了影像组学在多模态、小样本等场景下的分析效果；另一方面，为相关疾病的智能分析提供理论与技术支撑。