基于生成对抗网络的细胞形变动态分类

提出了一种基于生成对抗网络的细胞形变动态分类方法，以活细胞视频中的细胞形变动态为对象，引入分类器辅助的生成对抗网络结构同步训练生成对抗网络和分类网络，通过生成对抗网络产生的数据提高了原本分类网络分辨细胞形变动态的性能.首先，细胞动态图像被用于将活细胞视频中的时间维度进行压缩，使其从视频域映射到图像域以方便生成对抗网络的构建.其次，基于分类器辅助的生成对抗网络结构，将分类网络的分类信息作为辅助信息来改善生成对抗网络对多类样本的生成，同时生成网络生成的多类样本可以反过来优化分类网络对于细胞动态形变的分类性能.在构建的活细胞视频数据库上，可以验证提出方法能有效地捕获细胞视频中的空时细胞形变动态，并且其分类的性能优于其它主流方法.

Classification of Cell Deformation Dynamics Based on Generative Adversarial Networks

A generative adversarial networks (GANs) based model was proposed to classify cell deformation dynamics. In the framework, an auxiliary classifier GANs (AC-GANs) were introduced to simultaneously train GANs and a classification network for cell deformation dynamics in live-cell videos. The generated samples from GANs could further enhance the performance of the original classification network. To facilitate application of GANs, cell dynamic image was used to encapsulate the cell dynamics in videos along the temporal dimension, making the cell dynamics information mapped from video area to image area for the construction of the GANs. Then, the classification information was employed in AC-GANs to improve the generation of multi-class samples for GANs, and these multi-class samples could enhance the performance of classification net for improving the cell dynamic deformation. Experimental results demonstrate that the proposed pipeline can effectively capture the spatio-temporal cell dynamics from the raw live-cell videos and outperforms existing methods on the live-cell database.