基于深度学习的医学图像分析域自适应研究

深度学习技术的广泛应用有力推动了医学图像分析领域的发展，然而大多数深度学习方法通常假设训练集和测试集是独立同分布的，这个假设在模型临床部署时很难保证实现，因此常出现模型性能下降、场景泛化能力不强的困境。基于深度学习的域自适应技术是提升模型迁移能力的主流方法，其目的是使在一个数据集上训练的模型，能够在另一个没有或只有少量标签的数据集上也获得较好结果。由于医学图像存在着样本获取和标注困难、图像性质特殊、模态差异等情况，这给域自适应技术带来很多现实挑战。首先将介绍域自适应的定义及面临的主要挑战，进而从技术角度分类总结了近年来的相关算法，并对比分析其优缺点；然后详细介绍了域自适应常用的医学图像数据集以及相关算法结果情况；最后，从发展瓶颈、技术手段、交叉领域等方面，展望了面向医学图像分析的域自适应的未来研究方向。

Survey of medical image analysis domain adaptation based on deep learning

The wide application of deep learning techniques has strongly promoted the development of the medical image analysis field. However, most deep learning methods usually assume that the training and test sets are independently and identically distributed, and this assumption is problematic to guarantee when the models are deployed clinically. Hence, the dilemma of model performance degradation and poor scene generalization ability often occurs. Deep learning-based domain adaptation techniques are the mainstream methods for improving model migration ability. It aims to enable the model trained on one dataset to obtain better results on another dataset with no or only a small number of labels. Due to the difficulties in sample acquisition and labelling, unique image properties and modal differences in medical images bring many practical challenges to domain adaptive technology. This paper firstly introduced the definition and primary challenges of the domain adaptation and then classified and summarized related algorithms in recent years from a technical point of view, compared and analyzed their advantages and disadvantages, and then introduced the medical image datasets commonly used in domain adaptation and related algorithm results in detail. Finally, this paper prospected the future research direction of domain adaptation for medical image analysis regarding development bottlenecks, technical means, and cross-cutting areas.