Microscopy image reconstruction with physics-informed denoising diffusion probabilistic model.

Area of Science:

• Biomedical imaging
• Computational microscopy
• Deep learning applications

Background:

• Light microscopy is crucial for biomedical research but limited by diffraction and optical imperfections.
• Deep learning methods are emerging to improve light microscopy but often produce artifacts.

Purpose of the Study:

• To develop a deep learning approach for enhancing light microscopy image quality.
• To address artifacts and hallucinations in deep learning-based image reconstruction.

Main Methods:

• A conditioned diffusion model in a physics-informed neural network architecture was employed.
• The model's loss function incorporates the physics of light propagation in microscopy.
• Synthetic datasets were utilized for training to overcome data limitations.

Main Results:

• Consistent enhancements in image quality were observed.
• Substantial reductions in image artifacts and hallucinations were achieved.
• The method outperformed state-of-the-art techniques in image reconstruction.

Conclusions:

• The developed physics-informed deep learning technique offers an accessible way to obtain higher quality microscopy images.
• This approach holds significant potential for advancing biomedical studies and diagnostics.