近年来，医学人工智能领域发展方兴未艾，而影像技术为其提供了大量直观而可信的信息。然而，医学影像多存在多源异构、维度高、样本少、无结构等特性，造成分析结果鲁棒性欠佳。本项目拟以基于大脑影像的脑疾病早期诊断为问题驱动，通过对影像多尺度特征表达与融合和基于脑启发的小样本学习两大核心问题研究，为脑疾病早期诊断提供理论与方法论上的支撑。具体地，首先，将临床特征理解通过正则约束引入特征表达模型，从而提高模型的正则性和特征表达的鲁棒性。进一步，通过建立多尺度特征融合的有效机制解决大脑影像特征的多源异构性；其次，通过非线性度量学习建立基于数据驱动的数据分布最优描述，从而减少数据偏见和后续数据分类难度；再次，基于数据分布的最优描述，建立基于特征迁移学习和迁移过程学习的小样本学习机制，从而实现小样本数据的精确分类。最终，形成大脑影像智能分析的新理论和新方法，同时也为从小样本学习角度研究类脑智能提供有益探索。

In recent years, AI in medicine has a great development, while imaging techniques offer amounts of intuitive and believable information. However, medical images always have the property of multi-source heterogeneity, huge dimension, small sample and non-structure, and hence they lead to the lack of robustness for the medical image analysis. Therefore, in this project, we will study two essential questions about robust feature representation and small sample learning for image-based early diagnosis for brain disease. Concretely, we will first introduce the clinical knowledge as a regular constraint to the model to improve its regularity and robustness of feature representation. Further, by establishing the effective mechanism for multi-scale features fusion, we solve the multi-source heterogeneity of images. Second, by designing the nonlinear metric learning framework, we will establish the best description of the data distribution to reduce the bias of the data and difficulty of data classification. Thirdly, under the best description of data distribution, we will establish the effective mechanisms of feature transfer and transfer process learning to realize the accurate classification of small samples. Finally, we will form a novel theory and methodology for intelligent analysis of brain images, as well as offer a useful exploration for researches of brain-like intelligence from the viewpoint of small sample learning.