眼底图像分析中的迁移学习算法和应用研究

In recent years, fundus image analysis technologies have made rapid progress. However, the performance of most existing fundus image analysis models is still a bit far from satisfaction. The main reason is due to lack of labeled fundus images as well as training and test images coming from different sources. In this project, we propose to introduce transfer learning to resolve the above two issues. More specifically, we focus on three scenarios of transfer learning: i) instances lie in the same feature space but follow a different data distribution; ii) instances lie in different feature spaces; and iii) no source instances are available but only source models. By studying those three scenarios, we aim at proposing a generalized transfer learning framework that is able to reduce the data distribution mismatch between different domains, as well as to handle negative transfer through effective selection of source data or source models. We will also utilize abundant unlabeled target training instances, which also helps our proposed framework learn robust models. Finally, we will apply our framework to solve several practical problems existing in fundus image analysis and relating to transfer learning. We hope to achieve pioneering research results, fuel the technological development of automatic computer-aided diagnosis, and make contributions for common people to provide them with effective and convenient ways of ocular disease screening and diagnosis in the future.

近年来，眼底图像分析技术得到了迅猛发展。然而，绝大部分现有的眼底图像分析模型在实际应用中的效果还是不尽如人意。这主要是因为带标签的眼底图像不足、训练图像和测试图像来源不同等原因造成的。在本项目中，我们提出引入迁移学习来有效解决上述问题。具体的讲，我们聚焦迁移学习的三个基本情况：i)数据处于同一特征空间但服从不同分布；ii)数据处于不同特征空间；iii)仅有源模型但无源数据。通过研究和总结，我们致力于提出统一的迁移学习框架来减少不同领域数据分布的差异性，并通过合理选择源数据或源模型来有效避免负迁移的影响，同时引入大量的未标注目标训练样本来进一步充实我们的理论框架。最后，我们以眼部图像分析中的涉及到迁移学习的若干实际问题作为应用落脚点。我们希望通过本项目做出具有开拓性的科研成果，进一步推动眼部疾病的全自动计算机辅助诊断技术的发展，为将来给普通民众提供有效的、便利的眼睛疾病普查服务做出贡献。

本项目针对训练图像和测试图像来源不同、带标签的眼底图像不足等问题，分别研究了跨域数据的青光眼眼底图像数据分割和弱监督的迁移学习算法。1）对青光眼的眼底图像中视杯视盘的图像分割问题，关注视杯视盘的联合分割，取得以下研究进展：i）我们将视杯视盘的分割问题转化为视杯视盘的检测问题，视杯视盘的分割结果可以通过求最小边界框的内切椭圆得到。ii）提出了一个称作JointRCNN的端到端的神经网络模型，在网络中运用了注意力机制以提升分割效果；2）对弱监督的迁移学习。我们提出了一种针对目标定位的基于弱监督学习的迁移学习框架。此外，利用易于收集的目标图像的图像级标注来提高跨域分割的性能，构建了基于弱监督学习的多层级课程迁移学习框架。最后，我们开发了一套多病种的眼底图像数据库，对这些眼底彩色照片对规范化管理，为眼底疾病的辅助诊疗研究的数据获取和使用提供方便。