Deep learning-enhanced fluorescence microscopy via confocal physical imaging model.

Area of Science:

• Microscopy and Imaging Science
• Computational Biology
• Optical Physics

Background:

• Confocal microscopy is vital for high-resolution imaging in biology and industry.
• Deep learning aids micrograph reconstruction but often ignores imaging physics, causing aliasing issues.
• Existing methods struggle with multi-scale image pair alignment and generalization.

Purpose of the Study:

• To develop a deep learning method for confocal microscopy image reconstruction that addresses aliasing and improves fidelity.
• To integrate an image degradation model based on physical principles into deep learning for enhanced performance.
• To eliminate the need for precise image alignment in network training.

Main Methods:

• Developed an image degradation model using Richards-Wolf vectorial diffraction integral and confocal imaging theory.
• Generated low-resolution training images by degrading high-resolution counterparts using the physical model.
• Combined a residual neural network with a lightweight feature attention module and the degradation model.

Main Results:

• Achieved high fidelity and generalization in confocal image reconstruction.
• Outperformed non-negative least squares and Richardson-Lucy deconvolution algorithms.
• Reached a structural similarity index above 0.82 and improved peak signal-to-noise ratio by over 0.6 dB.
• Demonstrated broad applicability across different deep learning networks.

Conclusions:

• The proposed physically-informed deep learning model significantly enhances confocal microscopy image reconstruction quality.
• This approach mitigates aliasing problems and improves generalization compared to conventional methods.
• The method shows promise for various applications requiring high-fidelity microscopic imaging.