Sytwu Katherine, Groschner Catherine, Scott Mary C
Molecular Foundry, Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, CA 94720, USA.
Materials Science and Engineering, University of California Berkeley, Berkeley, CA 94720, USA.
Microsc Microanal. 2022 Sep 13:1-9. doi: 10.1017/S1431927622012466.
Trained neural networks are promising tools to analyze the ever-increasing amount of scientific image data, but it is unclear how to best customize these networks for the unique features in transmission electron micrographs. Here, we systematically examine how neural network architecture choices affect how neural networks segment, or pixel-wise separate, crystalline nanoparticles from amorphous background in transmission electron microscopy (TEM) images. We focus on decoupling the influence of receptive field, or the area of the input image that contributes to the output decision, from network complexity, which dictates the number of trainable parameters. For low-resolution TEM images which rely on amplitude contrast to distinguish nanoparticles from background, we find that the receptive field does not significantly influence segmentation performance. On the other hand, for high-resolution TEM images which rely on both amplitude and phase-contrast changes to identify nanoparticles, receptive field is an important parameter for increased performance, especially in images with minimal amplitude contrast. Rather than depending on atom or nanoparticle size, the ideal receptive field seems to be inversely correlated to the degree of nanoparticle contrast in the image. Our results provide insight and guidance as to how to adapt neural networks for applications with TEM datasets.
经过训练的神经网络是分析日益增长的科学图像数据的有前途的工具，但目前尚不清楚如何针对透射电子显微镜图像的独特特征对这些网络进行最佳定制。在这里，我们系统地研究了神经网络架构的选择如何影响神经网络在透射电子显微镜（TEM）图像中从非晶背景中分割出结晶纳米颗粒（即逐像素分离）的方式。我们专注于将感受野（即对输出决策有贡献的输入图像区域）的影响与网络复杂度（决定可训练参数数量）的影响分离开来。对于依赖幅度对比度来区分纳米颗粒与背景的低分辨率TEM图像，我们发现感受野对分割性能没有显著影响。另一方面，对于依赖幅度和相位对比度变化来识别纳米颗粒的高分辨率TEM图像，感受野是提高性能的重要参数，尤其是在幅度对比度最小的图像中。理想的感受野似乎与图像中纳米颗粒对比度的程度呈反比，而不是取决于原子或纳米颗粒的大小。我们的结果为如何使神经网络适用于TEM数据集的应用提供了见解和指导。
