基于生成对抗网络的医学数据域适应研究

Study on domain adaptation of medical data based on generative adversarial network

在医疗影像辅助诊断研究中,研究者通常使用不同医院(多域)的数据,但当其中一个域的训练样本较少时,模型在该域的测试集上的分类结果将会很差。针对此问题,提出一种基于生成对抗网络的分类方法进行男女脑影像差异的域适应研究,首先使用生成对抗网络学习不同域的数据分布,并提取关键特征,然后基于提取的关键特征研究不同域的男女脑影像差异。实验表明,该方法在仅有少量数据参与训练的域上也能取得80%以上的分类准确度。

In the study of medical imaging aided diagnosis,researchers often collect a lot of training data coming from different hospitals (named variety fields).But because of the certain field has insufficient training data,the deep learning model would get very poor performance on the test data of this field.To mitigate this problem,a method to study domain adaptation of the difference between male and female brain images based on the generative adversarial network was proposed.The data distribution of different domains was learned and the key features were extracted by using the generative adversarial network,and then the differences between male and female brain images in different domains were studied based on the extracted key features.Experiments show that the method can also achieve more than 80% recognition accuracy in the domain with only a small amount of data involved in training.