Virtual organelle self-coding for fluorescence imaging via adversarial learning.

Area of Science:

• Cell biology
• Biophysics
• Computational biology

Background:

• Traditional fluorescence microscopy requires extensive sample preparation, including chemical fixation and staining, which is time-consuming and costly.
• Deep learning offers a potential solution to streamline imaging processes by enabling virtual generation of fluorescence data.

Purpose of the Study:

• To evaluate the efficacy of deep learning methods in predicting fluorescence images based on structural and functional relationships.
• To introduce a novel virtual fluorescence staining method (VirFluoNet) for generating subcellular compartment-specific molecular fluorescence labels.

Main Methods:

• Developed VirFluoNet, a virtual-fluorescence-staining method utilizing deep neural networks, specifically conditional generative adversarial networks.
• Trained the model on microscopy datasets from MDA-MB-231 (breast cancer) and U2OS (bone-osteosarcoma) cell lines.
• Assessed performance using metrics like Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and a novel tolerance level metric.

Main Results:

• For MDA-MB-231 cells, F-actin signal input outperformed phase-contrast for vinculin prediction.
• U2OS cell predictions achieved satisfactory performance metrics compared to ground truth (e.g., MAE <0.005, PSNR >40 dB, SSIM >0.925 for specific markers).
• Successfully predicted fluorescence labels for 4',6-diamidino-2-phenylindole/Hoechst, endoplasmic reticulum, and mitochondria.

Conclusions:

• Deep learning image-regression is effective for predicting fluorescence microscopy data when input and output labels share structural or functional relationships.
• The VirFluoNet approach shows promise for modeling intracellular spatial relationships and detecting alterations in living cells.
• This method can reduce the reliance on physical sample preparation for fluorescence imaging.