Deep learning-based photodamage reduction on harmonic generation microscope at low-level optical power.

Area of Science:

• Optical bioimaging
• Microscopy
• Artificial intelligence

Background:

• High-quality imaging in optical bioimaging often conflicts with cellular health due to phototoxicity.
• Harmonic generation microscopy (HGM), including second harmonic generation (SHG) and third harmonic generation (THG), faces this trade-off.
• Minimizing light exposure is critical for preserving cell viability during imaging.

Purpose of the Study:

• To develop a deep-learning-based power-enhancement (PE) model for HGM.
• To enable the prediction of high-power HGM images from low-power inputs.
• To reduce phototoxicity and photodamage in HGM while maintaining image quality.

Main Methods:

• Implementation of a deep-learning-based power-enhancement (PE) model within a harmonic generation microscope (HGM).
• Training the PE model using low-power HGM images to predict corresponding high-power images.
• Testing the model's efficacy on both normal and abnormal skin tissue data.

Main Results:

• The PE model successfully predicted high-power HGM images from low-power inputs.
• The model significantly reduced the risk of phototoxicity and photodamage.
• The PE model, trained on normal skin, accurately predicted abnormal skin data, aiding cancer cell identification.

Conclusions:

• Deep-learning-based power enhancement offers a solution to the high-quality imaging versus cellular health trade-off in HGM.
• The PE model demonstrates utility in dermatopathology for identifying cancerous cells.
• The PE model holds promise for both in-vivo and ex-vivo HGM applications.