Virtual organelle self-coding for fluorescence imaging via adversarial learning

Our study introduces an application of deep learning to virtually generate fluorescence images to reduce the burdens of cost and time from considerable effort in sample preparation related to chemical fixation and staining.

The objective of our work was to determine how successfully deep learning methods perform on fluorescence prediction that depends on structural and/or a functional relationship between input labels and output labels.

We present a virtual-fluorescence-staining method based on deep neural networks (VirFluoNet) to transform co-registered images of cells into subcellular compartment-specific molecular fluorescence labels in the same field-of-view. An algorithm based on conditional generative adversarial networks was developed and trained on microscopy datasets from breast-cancer and bone-osteosarcoma cell lines: MDA-MB-231 and U2OS, respectively. Several established performance metrics-the mean absolute error (MAE), peak-signal-to-noise ratio (PSNR), and structural-similarity-index (SSIM)-as well as a novel performance metric, the tolerance level, were measured and compared for the same algorithm and input data.

For the MDA-MB-231 cells, F-actin signal performed the fluorescent antibody staining of vinculin prediction better than phase-contrast as an input. For the U2OS cells, satisfactory metrics of performance were archieved in comparison with ground truth. MAE is <0.005, 0.017, 0.012; PSNR is >40  /  34  /  33  dB; and SSIM is >0.925  /  0.926  /  0.925 for 4',6-diamidino-2-phenylindole/hoechst, endoplasmic reticulum, and mitochondria prediction, respectively, from channels of nucleoli and cytoplasmic RNA, Golgi plasma membrane, and F-actin.

These findings contribute to the understanding of the utility and limitations of deep learning image-regression to predict fluorescence microscopy datasets of biological cells. We infer that predicted image labels must have either a structural and/or a functional relationship to input labels. Furthermore, the approach introduced here holds promise for modeling the internal spatial relationships between organelles and biomolecules within living cells, leading to detection and quantification of alterations from a standard training dataset.