Patch-Based U-Net Model for Isotropic Quantitative Differential Phase Contrast Imaging.

Area of Science:

• Microscopy
• Image Processing
• Computational Biology

Background:

• Quantitative differential phase-contrast (qDPC) imaging is a label-free technique for visualizing weak phase objects.
• Traditional qDPC requires multiple intensity measurements for isotropic phase reconstruction, which is time-consuming.
• Fewer measurements in qDPC lead to anisotropic phase distribution and loss of spatial frequency information.

Purpose of the Study:

• To investigate the feasibility of using deep learning (DL) for isotropic qDPC microscopy with minimal measurements.
• To develop a DL model capable of generating isotropic phase images from anisotropic inputs.
• To improve the efficiency and accuracy of qDPC imaging for biological studies.

Main Methods:

• A U-net convolutional neural network architecture was employed for image reconstruction.
• The U-net model was trained using a patch-wise approach with seven types of living cell images.
• The model was trained to generate 12-axis isotropic reconstructed images from 1-axis anisotropic input images.

Main Results:

• The DL-based method successfully generated isotropic qDPC images with accuracy comparable to 12-axis measurements.
• Quantitative phase values were recovered within 66% to 97% of ground-truth values.
• The U-net model demonstrated superior performance (higher PSNR and SSIM) compared to CycleGANs for isotropic qDPC microscopy.

Conclusions:

• Deep learning, specifically the U-net architecture, offers a feasible and efficient solution for isotropic qDPC microscopy using limited measurements.
• The proposed method enhances phase uniformity and retrieves missing spatial frequencies in reconstructed images.
• This DL-based approach has the potential to advance high-resolution quantitative studies in cell biology.