DeepDOF-SE: affordable deep-learning microscopy platform for slide-free histology

Paper pdf will be included soon...

EDOF network

Required packages

The required packages can be found in the deepdof-se.yml file. One can also directly use the yml file to create a virtual environment with all the packages. To do so with Anaconda, see this tutorial

EDOF dataset

The training, validation, and testing set used for the EDOF network can be found at https://zenodo.org/record/3922596

One network EDOF

Uses a single U-net to deconvolve the coded blurred image. Faster to train than the dual U-net version. To use the code, change the file paths at the beginning of each file. Contains the following files:

• MuseEDOF_cubic_RGB_sep_step1.py: trains the U-net with a cubic phase mask
• MuseEDOF_cubic_RGB_sep_step2.py: trains both the U-net and the phase mask jointly
• Network_RGB.py: U-net
• recon_RGB.py: Reconstruct captured coded-blurred image after the network is trained and fine-tuned
• a_zernike_cubic_150mm.mat: contains the coefficient of the cubic mask
• zernike_basis_150mm.mat: contains the Zernike basis of the mask

Dual network EDOF

Uses a U-net for each fluorescence dye channel. Higher reconstruction quality. To use the code, change the file paths at the beginning of each file. Contains the following files:

• CM_MicroDualEDOF_cubic_rms_dualimage_dualunet_128x21_2step_step1.py: trains dual U-net with a cubic phase mask
• CM_MicroDualEDOF_cubic_rms_dualimage_dualunet_128x21_2step_step2.py: trains both the U-net and the phase mask jointly
• dualunet_reconstruct.py: Reconstruct captured coded-blurred image after the network is trained and fine-tuned
• Network_c1.py: 1 of the 2 U-net
• Network_c2.py: 1 of the 2 U-net
• a_zernike_cubic_150mm.mat: contains the coefficient of the cubic mask
• zernike_basis_150mm.mat: contains the Zernike basis of the mask

Virtual staining network

The CycleGAN virtual staining network is based on the Tensorflow implementation by Harry Yang link. Contains the following files:

• cyclegan_step1.ipynb: step 1 of the cycleGAN training. Assumes the data has paired images (fluorescence & Beer-Lambert virtual staining)
• cyclegan_step2.ipynb: step 2 of the cycleGAN training. Assumes the data has unpaired images (fluorescence & FFPE H&E)
• image_preprocess.py: contains utility functions that preprocess the images (e.g. normalization)
• network_losses.py: contains the functions that compute different loss terms (e.g. identity loss)
• resnet_network.py: contains the resnet class
• testcode_folder.ipynb: code that performs virtual staining after the network is trained.

Virtual staining dataset

This data set contains patient data and is available upon reasonable request. Please contact the corresponding authors Ashok Veeraraghavan or Rebecca Richards-Kortum.