智能环境下机器学习在解决跨模态医疗知识推理中体现出显著优势,但也存在样本匮乏、标注成本高及模型嵌入后核心属性获取困难问题。为此在多尺度上构建基于SVM的多模态数据特征分析模型,实现面向开放域的跨模态数据模式切分、特征抽取及碎片化医疗知识非线性融合;构建基于迁移学习与弱监督学习的医学影像增强模型,解决专家标注昂贵且病例稀少造成深度学习训练无法完成问题;对3D医学图像建模,实现图像互补信息和空间上下文信息融合,提高像素级分类/分割精度;构建基于深度学习的跨模态医疗知识推理模型集,将深度学习与关系网络模块结合,捕获关系推理核心共同属性,提高知识推理对小任务和输入变量处理的鲁棒性;提出一种自适应的卷积神经网络,解决大规模知识库/知识图谱推理中,路径排序算法需在离散空间进行知识表示,难以对相似实体-关系进行比较的问题;在实验平台上完成模型集验证,提升医疗决策能力,为构建医学知识图谱提供理论依据。

In ambient intelligence, machine learning shows significant advantages in cross-modal medical knowledge reasoning. However, it also has some defects on model training, such as sample missing and high cost labeling. Meanwhile, machine learning is difficult to obtain core attributes after model embedded. For these reasons, a multi-modal feature selecting model based on SVM is proposed on multi-granularity to realize cross-modal pattern segmentation, feature extraction and nonlinear integration of fragmented medical knowledge in open domain. Based on transfer learning and weak supervised learning, a medical image enhancement model is implemented to provide solutions on uncompleted deep learning training due to expensive expert labeling and lack of case data. A 3D medical image model is constructed to realize the fusion of image complementary information and spatial context information, that also can improve the accuracy of classification/segmentation on pixel level. A set of cross-modal knowledge reasoning models is put forward on the base of deep learning, where relation networks (RNs) is used as a simple plug-and-play module to captures the core common properties of relational reasoning. As a result, the robustness of processing on small tasks and input variables can be improved. Because Path-Ranking Algorithm(PRA) need to run in fully discrete space, an adaptive Convolution Neural Network(CNN) is developed to evaluate and compare similar entities and their relations in a large scale knowledge base/graphs. On the experimental platform equipped with intelligent patient robots, the set of validation of knowledge reasoning models will be completed. As the conclusion of the research, we hope to greatly improved the ability on medical decision-making. And research findings also can provide scientific basis and theoretical guidance for the construction of medical knowledge graphs.