Deep learning enhanced quantitative phase imaging in digital holographic microscopy with attention empowered deep...

Area of Science:

• Optics and Photonics
• Computational Imaging
• Biomedical Engineering

Background:

• Digital holographic microscopy (DHM) enables quantitative phase imaging (QPI).
• Pixelated polarization cameras in DHM offer snapshot QPI but reduce resolution and introduce coherent noise.
• Existing methods struggle to simultaneously address resolution loss and noise in QPI.

Purpose of the Study:

• To develop an advanced deep image prior (DIP) method for simultaneous phase image interpolation and denoising.
• To overcome the limitations of supervised learning by utilizing an unsupervised approach for QPI enhancement.
• To improve the quality of quantitative phase images obtained from DHM with pixelated polarization cameras.

Main Methods:

• Proposed an attention-enhanced double Unet (ADUnet+) model for unsupervised learning.
• Implemented a deep image prior (DIP) strategy, leveraging the implicit priors of the ADUnet+ architecture.
• Employed an attention mechanism within ADUnet+ to effectively extract underlying phase map features.
• Compared different learning strategies, identifying a sequential two-step mapping as superior to a one-step procedure.

Main Results:

• The proposed DIP method successfully performs simultaneous phase image interpolation and denoising.
• The attention mechanism in ADUnet+ effectively rescales feature map weights, enhancing feature extraction.
• A sequential two-step mapping learning strategy demonstrated superior performance.
• The method significantly outperformed baseline methods in both simulated and experimental QPI data.

Conclusions:

• The developed ADUnet+-based DIP method effectively enhances quantitative phase images from DHM.
• The unsupervised approach eliminates the need for large training datasets, offering a practical solution.
• The method demonstrates high universal capability and robustness across various sample types.
• This technique offers a significant advancement for high-quality, noise-free quantitative phase imaging in microscopy.