Generating Cytoplasmic Fluorescence Image from a Single Transmitted Light Microscopy Image with Pyramid Pix2pix

Jianyuan Zeng, Tianzhao Zhong, Jincai Huang, Jinfeng Peng, Shaojun Liu

Abstract: Obtaining cell fluorescence microscopy images is a time-consuming and resource-intensive task. Additionally, the fluorescence dyes can cause damage to cells, thereby affecting the final judgment. Alternatively, fluorescence microscopy images from transmitted light microscopy images can be generated using machine learning methods. In this paper, we employed CycleGAN and Pix2pix pyramid generative model to achieve this goal. We introduced data preprocessing to tackle the dataset’s challenges where the image amount, size, and channels vary a lot. We trained models for different modalities of fluorescence images to reduce the impact of modalities on the model training process. Ultimately, our approach achieved promising results in terms of structural similarity, mean absolute error, Peasorn’s correlation coefficient, Euclidean distance, and Cosine distance. The code is available at https://github.com/achuansztu/pyramidPix2pixfor-lightmycell/tree/main.