Regularized Gradient Statistics Improve Generative Deep Learning Models of Super Resolution Microscopy.

Area of Science:

• Microscopy
• Deep Learning
• Image Processing

Background:

• Super-resolution microscopy aims to overcome the diffraction limit of light.
• Deep learning models are increasingly used for image reconstruction in microscopy.
• Current models may struggle with capturing fine details and generalizing across different image types.

Purpose of the Study:

• To improve the quality of super-resolution fluorescence microscopy images generated by deep learning models.
• To introduce a novel regularization technique based on natural-scene image statistics.
• To evaluate the effectiveness of this regularization on a state-of-the-art generative model.

Main Methods:

• Regularizing training data's gradient and Laplacian statistics to mimic natural scenes.
• Utilizing the BioSR dataset of paired diffraction-limited and super-resolution images.
• Evaluating the method with a Conditional Variational Diffusion Model (CVDM).

Main Results:

• The proposed regularization technique enhances visual detail in generated super-resolution images.
• Images produced using the new prior exhibit improved small-scale structure.
• The regularization improved the generalization power of the CVDM model.

Conclusions:

• Regularizing signal gradient statistics is an effective method for improving deep learning-based super-resolution microscopy.
• This preprocessing approach is compatible with various supervised machine learning models.
• The technique shows particular promise for images of filamentous structures where natural-scene priors are applicable.