基于生成对抗网络的乳腺癌组织病理图像样本均衡化

Generative adversative nets based sample equalization of breast tumor tissue slice

针对乳腺癌组织病理图像中各类样本的不均衡性,本研究构建了基于条件控制和深度卷积的两种乳腺癌组织病理图像样本生成对抗网络并生成人造样本,以实现样本均衡化。实验验证了两种网络用于数据均衡化的可靠性,发现在处理不均衡数据时基于深度卷积的生成对抗网络效果更好,基于条件控制的生成对抗网络鲁棒性更强;分析了增加人造样本对深度学习算法的影响。结果表明,本研究提出的数据均衡化方法将训练所得网络的分类准确率平均提升了近5%,生成对抗网络可以缓解深度学习在医学领域的应用中数据分布不平衡的问题。

Aiming at the problem of unbalanced cell data and low classification accuracy of breast tumor tissue sections,we constructed two kinds of mammary cancer cell sample generation antagonism networks based on conditional control and deep convolution to generate artificial samples and accomplish sample equalization.Experiments verified the reliability of the two networks for data equalization.It was found that the effect of conditional deep convolutional generative adversarial networks was better and the robustness of conditional generative adversarial nets was stronger when dealing with unbalanced data.The influence of adding artificial samples on deep learning algorithm is analyzed.The results show that the proposed data equalization method improves the classification accuracy of the trained network by 5%,the generative adversarial nets can alleviate the problem caused by imbalance data in the application of deep learning in the medical field.