Fourier ptychographic microscopy image enhancement with bi-modal deep learning.

Area of Science:

• Digital pathology
• Microscopy
• Artificial intelligence in diagnostics

Background:

• Whole slide imaging and Fourier ptychographic microscopy (FPM) offer advanced capabilities for automated disease detection.
• Challenges arise with thick samples where depth of field issues cause contrast changes in phase images, limiting diagnostic accuracy.
• Existing methods struggle to maintain sub-cellular contrast and focus precision in high-resolution microscopy.

Purpose of the Study:

• To develop a robust method for enhancing sub-cellular compartment contrast in FPM images, particularly for thick biological samples.
• To improve the usability of phase images from FPM by compensating for focus variations and contrast changes.
• To demonstrate the efficacy of the developed method in improving disease detection sensitivity without compromising specificity.

Main Methods:

• A bi-modal U-Net artificial neural network was trained using both intensity and phase images from FPM.
• A reference database was constructed using FPM reconstruction algorithms and virtual Z-stacking to select optimal focal planes.
• The U-Net was trained to simultaneously enhance targeted sub-cellular compartment contrast and correct for focus imprecision.

Main Results:

• The trained U-Net effectively reinforced sub-cellular compartment visibility in both intensity and phase images.
• The method demonstrated robustness across large fields of view at high resolution.
• In the use-case of *Plasmodium falciparum* detection, improved detection sensitivity was achieved without a loss in specificity.

Conclusions:

• The bi-modal U-Net approach offers a significant advancement in post-reconstruction image processing for FPM.
• This method is generalizable and applicable to various demanding biological screening applications requiring high-resolution imaging.
• The technique enhances automated diagnosis by improving image quality and compensating for optical limitations in microscopy.