Self-supervised learning of hologram reconstruction using physics consistency

Luzhe Huang, Hanlong Chen, Tairan Liu, Aydogan Ozcan, Self-supervised learning of hologram reconstruction using physics consistency. Nat Mach Intell 5, 895–907 (2023).

[Paper] [Video] [Citation]

Environment requirements

The codes was tested on Windows 11 and Linux, with Python and PyTorch. Packages required to reproduce the results can be found in requirements.txt. The following software / hardware is tested and recommended:

• Python >= 3.9
• CUDA 11.2
• cuDNN >= 8.0
• Intel i9-12900F
• Nvidia RTX 3090
• RAM >= 64 GB

File structure

This repository contains codes for GedankenNet and GedankenNet-Phase, and demo models and data for each network.

GedankenNet
|   README.md
|   requirements.txt
|
|---GedankenNet
|   |   generate_random_image_parallel.py
|   |   train_Gedanken.py
|   |   test.py
|   |   ...
|
|---GedankenNet_Phase
|   |   train_GedankenP.py
|   |   testP.py
|   |   ...
|
|---Models
|
|---demo_data
|   |   stained_tissue
|   |   unstained_tissue

/GedankenNet and /GedankenNet_Phase contain the codes to train and test the two models.

/Models include two demo models for GedankenNet and GedankenNet-Phase respectively.

/demo_data/stained_tissue include two demo FOVs of lung and Pap smears (complex object), corresponding to the results of /GedankenNet.

/demo_data/unstained_tissue include two demo FOVs of kidney tissue (phase-only object), corresponding to the results of /GedankenNet_Phase.

Test

To test the two demo models (download via Google Drive) and reproduce some results shown in the paper, following these steps:

• Download checkpoints in /Models and mat files in /demo_data
• Change TEST_PATH in test.py or testP.py to the path of corresponding mat files
• Run test.py or testP.py. The outputs will be saved in outputs folders

Train

To train GedankenNet and GedankenNet-Phase models from scratch, follow these steps:

• Run /GedankenNet/generate_random_image_parallel.py to generate 100K artificial images, and then split these images into train, valid and test sets
• Change TRAIN_PATH and VALID_PATH to the locations of the generated artificial image datasets in train_Gedanken.py or train_GedankenP.py
• Run train_Gedanken.py or train_GedankenP.py, checkpoints will be saved in Models folders

Alternatively, we have uploaded one artificial image dataset with 99.9K training and 100 testing / validation images. Download from Google Drive.