Learned spatially varying microscopy model with adaptive point spread functions.

Area of Science:

• Optical microscopy
• Computational imaging
• Biomedical engineering

Background:

• Lensless fluorescence microscopy systems exhibit spatial variability requiring precise calibration.
• Traditional calibration methods often involve complex, stepwise procedures using motorized stages.

Purpose of the Study:

• To introduce a novel framework, the learned spatially varying microscopy model (LSVMM), for accurate modeling of spatial variations in microscopy.
• To provide a practical and robust alternative to conventional calibration techniques.

Main Methods:

• LSVMM learns spatially varying point spread functions using randomly distributed bead images.
• An adaptive warp module supports the learning process.
• The framework is evaluated on lensless microscopes with inherent variability.

Main Results:

• LSVMM effectively captures position-dependent optical responses and achieves robust modeling accuracy with minimal preprocessing.
• The model demonstrates superior performance compared to conventional spatially varying models, enhancing median peak signal-to-noise ratio by 2-7 dB and median multi-scale structural similarity by 0.02-0.10.
• LSVMM functions as a differentiable forward model, enabling synthetic data generation and integration with deep learning pipelines.

Conclusions:

• LSVMM offers a practical and accurate solution for modeling spatial variability in lensless fluorescence microscopy.
• The framework's adaptability makes it suitable for various spatially varying imaging modalities.
• This approach simplifies calibration and enhances image reconstruction quality.