Unsupervised inter-domain transformation for virtually stained high-resolution mid-infrared photoacoustic microscopy...

Area of Science:

• Biomedical Imaging
• Optical Microscopy
• Computational Biology

Background:

• Mid-infrared photoacoustic microscopy (MIR-PAM) offers label-free biochemical information.
• MIR-PAM's spatial resolution is limited by long optical wavelengths compared to confocal fluorescence microscopy (CFM).
• Existing methods lack explainability and stability for image transformation.

Purpose of the Study:

• To develop an explainable deep learning (XDL) framework for transforming low-resolution MIR-PAM images into high-resolution, virtually stained images.
• To improve the spatial resolution and interpretability of MIR-PAM.
• To enable label-free, high-resolution cellular imaging.

Main Methods:

• An unsupervised generative adversarial network (GAN) was employed for inter-domain image transformation.
• A saliency constraint was integrated into the GAN for enhanced explainability.
• The XDL framework was validated on cultured human cardiac fibroblasts, comparing results with CFM images.

Main Results:

• The XDL framework successfully transformed low-resolution MIR-PAM images into high-resolution, confocal-like images.
• The method accurately identified cell nuclei and filamentous actins.
• The XDL framework demonstrated stable and reliable performance, ensuring similar saliency between image domains.

Conclusions:

• Explainable deep learning-based MIR-PAM (XDL-MIR-PAM) enables label-free, high-resolution duplexed cellular imaging.
• This technique overcomes the spatial resolution limitations of conventional MIR-PAM.
• XDL-MIR-PAM offers significant benefits for cell biology research by providing detailed, label-free imaging.