Self-Net

Self-Net is a deep-learning-based Python module for improving the resolution isotropy of volumetric fluorescence microscopy. This is an official CODE repository for a research article titled "Deep self-learning enables fast, high-fidelity isotropic resolution restoration for volumetric fluorescence microscopy".

Citations

Ning K, Lu B, Wang X, et al. Deep self-learning enables fast, high-fidelity isotropic resolution restoration for volumetric fluorescence microscopy[J]. Light: Science & Applications, 2023, 12(1): 204.

Software

• Python 3.9 (tested on)
• Conda
• Pytorch 1.13.1. (tested on)
• CUDA version 11.7 (tested on)
• PyCharm 2022.3
• Windows 10

Main Dependencies

• numpy

• scipy

• pytorch

• opencv

• scikit-image

• tifffile

Hardware

• CPU or GPU that supports CUDA CuDNN and Pytorch 1.13.1.
• We tested on NVIDIA GeForce RTX 3090 (24 GB) and TITAN Xp (12 GB).

Usage

1. For a given anisotropic 3D image stack (tiff format), first run the matlab code 'image_slice.m' to generate lateral image slices (in the xy folder), down-sampled image slices (in the xy_lr folder), and axial image slices (in the xz or yz folder).

Input parameters: raw_path, data_name, scale

2. Run the python code 'Generate_training_data.py' to generate file: 'train_data.npz'.

Input parameters: path, raw_data_path, train_data_path, signal_intensity_threshold, xy_interval, xz_interval

3. Run 'Train_Self_net.py' to train Self_Net.

Input parameters: path, min_v, max_v, imshow_interval

4. Run 'Self_net_output_volume.py' for isotropic restoration of the raw anisotropic image stack.

Input parameters: test_path, model_path, min_v, max_v, raw_img, scale