Generative Adversarial Networks for Augmenting Training Data of Microscopic Cell Images
Generative Adversarial Networks for Augmenting Training Data of Microscopic Cell Images
复制标题
DOI：
10.3389/fcomp.2019.00010
复制
发表时间：
2019-11-26
影响因子：
2.6
通讯作者：
Bretschneider, Till
中科院分区：
文献类型：
--
作者：
Baniukiewicz, Piotr;Lutton, E. Josiah;Bretschneider, Till
生成对抗网络（GAN）最近已成功用于创建逼真的2D合成显微镜细胞图像并预测中间细胞阶段。在本文中，我们强调GAN不仅可以用于创建针对不同荧光分子标记优化的合成细胞图像，而且通过使用GAN来增强涉及缩放或其他转换的训练数据，生物结构的固有长度尺度得到保留。此外，GAN还可以创建具有特定形状特征的合成细胞，例如，可以用于验证不同的特征提取方法。在这里，我们应用GANs来创建Dictyosteopathy细胞（ABD）细胞皮质中F-actin的荧光标记物的2D分布，膜受体（cAR 1）和皮质-膜连接蛋白（TalA）。最近更广泛地使用3D光片显微镜，其中获得足够的训练数据是相当困难的，比在2D，创造了显着的需求，新的方法来数据增强。我们证明了使用GAN直接生成合成3D细胞图像是可能的，但限制是训练时间过长，依赖于3D图像的高质量分割，并且在不重新训练网络的情况下无法自由调整z切片的数量。我们证明，在与细胞形状高度相关的分子标记的情况下，如我们示例中的F-actin，2D GAN可以有效地用于从单独生成的2D切片创建伪3D合成细胞数据。由于高质量的分段2D细胞数据更容易获得，这是使用效率较低的3D网络的有吸引力的替代方案。
Generative adversarial networks (GANs) have recently been successfully used to create realistic synthetic microscopy cell images in 2D and predict intermediate cell stages. In the current paper we highlight that GANs can not only be used for creating synthetic cell images optimized for different fluorescent molecular labels, but that by using GANs for augmentation of training data involving scaling or other transformations the inherent length scale of biological structures is retained. In addition, GANs make it possible to create synthetic cells with specific shape features, which can be used, for example, to validate different methods for feature extraction. Here, we apply GANs to create 2D distributions of fluorescent markers for F-actin in the cell cortex of Dictyostelium cells (ABD), a membrane receptor (cAR1), and a cortex-membrane linker protein (TalA). The recent more widespread use of 3D lightsheet microscopy, where obtaining sufficient training data is considerably more difficult than in 2D, creates significant demand for novel approaches to data augmentation. We show that it is possible to directly generate synthetic 3D cell images using GANs, but limitations are excessive training times, dependence on high-quality segmentations of 3D images, and that the number of z-slices cannot be freely adjusted without retraining the network. We demonstrate that in the case of molecular labels that are highly correlated with cell shape, like F-actin in our example, 2D GANs can be used efficiently to create pseudo-3D synthetic cell data from individually generated 2D slices. Because high quality segmented 2D cell data are more readily available, this is an attractive alternative to using less efficient 3D networks.