动态增量式的鲁棒多中心医疗影像数据分类方法研究

单一中心建立的医疗模型精度有限，使用多中心数据联合训练可以提高模型精度，但有两个难点：1）联合训练会降低模型的实时性和实用性并导致大量数据的集中，降低数据的安全性；2）模型精度容易受到不同医疗中心数据分布不一致问题和医疗数据类别不平衡问题的影响。为此，本项目提出基于在线（增量）学习的多中心研究新思路：仅根据新中心数据实时更新模型，提高模型的实时性、实用性和数据安全性。为进一步提高基于在线学习的多中心研究精度，研究内容包括：1）研究基于防遗忘的在线学习算法，确保模型在新旧中心数据上都能获得较高准确率，同时有能力学习更多不同医疗中心的数据；2）设计面向概念漂移的在线自适应优化算法，提高模型对多中心数据分布不一致问题的鲁棒性；3）设计面向数据不平衡的无参数在线学习算法，不依赖超参数就能缓解数据不平衡对多中心研究的影响。本项目的研究拓展了多中心研究方法，为促进医疗影像AI产品落地提供核心算法支持。

The model performance is limited when training on medical image data from single health center, while jointly training data from multiple centers can improve the model performance. However, there are two difficulties: 1) joint training reduces the real-time capability and practicability of model, and concentrates large amount of data. The data concentrating reduces data security; 2) the model performance easily suffers from the inconsistency problem across the medical images of multiple centers and data imbalance problem. Hence, this project proposes to learn multi-center medical image data using online (incremental) learning methods: the model is sequentially updated only based on new chunk of data in real-time. This study improves the real-time capability, practicability and data security. In order to further improve the classification performance of online learning based multi-center study, the research topics of this project include: 1) a new online learning method without catastrophic forgetting is designed, which ensures that the model can achieve high accuracy on data from both old and new health centers, and is equipped for learning data from more health centers; 2) to improve the robustness to the inconsistency across the images, we design a new adaptive online optimization method to overcome concept-drift problem; 3) a parameter-free loss function is designed for online imbalance learning, which resolves class imbalance problem in multi-center data without tuning any hyper-parameters. In conclusion, this project will provide methodologies for multi-center study and core algorithms for promoting the landing of AI-based medical image products.