基于迁移学习的脑图像分析及其应用研究

With development of imaging devices, it has been attracted wide attention for exploring the association between brain functional deficits and the underlying structural disruption related to brain disorders from neuroimaging data. Machine learning has becoming a new focus of research in brain imaging analysis since it can obtain the rules via automatically analyzing data and apply these rules to predict the unknown data. Based on the advanced research of transfer learning, in this project we will study several important problems in brain imaging analysis, including feature learning, classification and regression, and sample learning. Specifically, in this project we will 1) construct the heterogeneous multimodal features deep transfer learning model to select the effective feature subset and learning the combination of multimodal feature, thus avoiding the defects of existing multimodal methods which not to feature learning from the heterogeneous multimodal features; 2) develop a novel multi-domain transfer learning model, which can avoid the defects of existing multimodal transfer learning methods only deal with feature learning or classification; 3) design a novel multi-domain multi-label transfer learning model for feature learning, classification and regression, and sample learning, which can avoid the defects of existing multi-label learning methods only deal with single domain data; 4) apply the above-mentioned brain imaging analysis models and methods for early diagnosis of brain diseases. The study of this project will contribute to the theory and method of brain imaging analysis, and is also expected to achieve practical application results.

随着成像设备的发展和普及，从神经影像中探索脑功能障碍与脑疾病相关的结构性破坏之间的关联受到广泛关注。机器学习由于能够从数据中自动获得规律，并对未知数据进行预测，已成为脑疾病诊断中一个新的研究热点。本项目基于迁移学习的最新研究进展，对脑图像分析中的特征学习、样本学习和分类回归等重要问题进行研究。具体地，本项目将：1）构建基于异质特征空间的多模态深度迁移学习模型，用于寻找到有效的特征子集与多模态特征组合，克服现有方法不能从异质多模态数据进行特征学习的缺陷；2）提出多领域迁移学习模型，克服了传统迁移学习只针对特征学习或分类预测的弊端；3）设计多标记多领域迁移学习模型，克服了传统多标记学习只针对单一学习领域的弊端，构建了一个包含特征学习、样本学习以及分类回归预测的学习模型；4）将上述模型和方法应用于脑疾病早期诊断。通过本项目的研究不仅能为脑图像分析的理论与方法有所贡献，还望能取得实际的应用成果。

随着成像设备的发展和普及，从神经影像中探索脑功能障碍与脑疾病相关的结构性破坏之间的关联受到广泛关注。机器学习由于能够从数据中自动获得规律，并对未知数据进行预测，已成为脑疾病诊断中一个新的研究热点。本项目基于迁移学习的最新研究进展，对脑图像分析中的特征学习、分类回归等重要问题进行研究。具体地，本项目完成以下工作：1）提出多领域迁移学习模型，克服了传统迁移学习只针对特征学习或分类预测的弊端；2）设计鲁棒多标记多领域迁移学习模型，克服了传统多标记学习只针对单一学习领域的弊端，构建了一个包含特征学习以及分类预测的学习模型；3）设计并实现基于权值分布稀疏特征学习模型，克服了传统特征学习方法不能对最优特征子集权重排序问题。将上述模型和方法应用于早期阿尔茨海默病诊断，并取得了较好的诊断性能。