Author

Shirui, Guo

Description

Official PyTorch implementation of OCEM: A deep-learning Python module for reconstructing 3D volumes from encoded 2D measurements. Code for the paper "Off-axis Compressed Encoding Microscopy for High-Resolution, High-Throughput, High-Contrast Volumetric Imaging".

Software

• Python 3.11 (tested on)
• Conda
• PyTorch 2.3.0 (tested on)
• Windows 10 / 11
• PyCharm 2022.3+

Hardware

• CPU or GPU that supports CUDA, CuDNN, and PyTorch 2.3.0.
• Minimum requirement: 12 GB VRAM (e.g., TITAN). Recommended: 24 GB VRAM (e.g., NVIDIA GeForce RTX 3090 ).

Instructions

• Install PyTorch and other dependencies. The main Python dependencies include:

numpy
pytorch
tifffile
matplotlib
torchsummary

Usage

1. For a given image stack (tiff format), first run the python code Preprocess_main.py to process the raw datasets, generate the .npy format training data.

   • Input parameters: file_root, save_file, sample_number, channels, arrays, point_1 (crop coordinates), layers, speed

2. Verify or adjust the normalization coefficients and dataset loading parameters in data_loader_our.py.

   • Input parameters: layer, strip, speed, normalization scaling factors (self.input / self.output divisors)

3. Run train_our.py to train the deep learning reconstruction network (e.g., LW-UNet, ResUNet, or RCAN).

   • Input parameters: --train_save_file, --data_file, --data_type, --train_test_val, --channel_in, --channel_out

4. Run predict_our.py (or predict_our_oridata.py) for the reconstruction of the testing image stack. This utilizes the trained network to restore structural details across the axial depth.

   • Input parameters: file (test dataset path), file_write (output restoration path), model (weights load path), scale (normalization coefficient), patch_size, stride, speed, depth