OpenDeepE: Open-Source Deep-E / 2D Fully-Dense U-Net

Non-official Pytorch implementation of "Deep-E: a fully-dense neural network for improving the elevation resolution in linear-array-based photoacoustic tomography" (IEEE TMI)

Setup

Installation

git clone https://github.com/thu-bplab/OpenDeepE
cd OpenDeepE

Environment

Install requirements for OpenDeepE first.

pip install -r requirements.txt

Dataset Organization

Directory Structure

The dataset should be organized in the following hierarchical structure:

dataset_root/
├── human_vessel/           # Data category
    ├── train/              # Training set
    │   ├── sparse_image/   # Sparse image data
    │   └── gt_image/       # Ground truth images
    ├── val/                # Validation set
    │   ├── sparse_image/
    │   └── gt_image/
    └── test/               # Test set
        ├── sparse_image/
        └── gt_image/

File Naming Convention

All data files are in NIfTI format (.nii) with the following naming pattern:

{sample_index}_{slice_index}.nii

Naming Explanation:

• sample_index: 3D data sample number (e.g., 1, 2, 3...)
• slice_index: 2D slice index (e.g., 90, 91, 92...)

Examples:

• 1_90.nii - Slice 90 from 3D sample 1
• 1_91.nii - Slice 91 from 3D sample 1
• 1_92.nii - Slice 92 from 3D sample 1
• 2_90.nii - Slice 90 from 3D sample 2

Data Description

• sparse_image: Contains input sparse image data
• gt_image: Contains corresponding ground truth image data
• The code defaults to reading NIfTI format files

Usage Notes

Ensure that each sample has corresponding files with identical names and counts in both sparse_image and gt_image directories to maintain data alignment during training.

Example Structure

human_vessel/
├── train/
│   ├── sparse_image/
│   │   ├── 1_90.nii
│   │   ├── 1_91.nii
│   │   ├── 1_92.nii
│   │   └── ...
│   └── gt_image/
│       ├── 1_90.nii
│       ├── 1_91.nii
│       ├── 1_92.nii
│       └── ...
├── val/
│   ├── sparse_image/
│   └── gt_image/
└── test/
    ├── sparse_image/
    └── gt_image/

This organization ensures proper data handling and compatibility with the provided codebase.

Training

Tips

• The recommended PyTorch version is >=2.0. Code is developed and tested under PyTorch 2.0.1.
• If you encounter CUDA OOM issues, please try to reduce the batch_size in the training and inference configs.
• The details related to the original network are all in Model Detail. Each processing step is explained with comments, which corresponds to Figure 1 of the original paper.

Configuration

• Our sample training defaults to use 1 gpu with fp32 precision.
• You may modify the configuration to fit your own environment.

Run Training

• Please replace data related paths in the script file or train.py with your own paths and customize the training settings.
• An example training usage is as follows:

  # Example usage
  cd scripts
  bash train.sh
  

Inference on Trained Models

• You need to modify the $MODEL_PATH in the testing script.
• An example inference usage is as follows:

  bash test.sh
  

Acknowledgement

• We thank the authors of the original paper for their great work!
• This project is supported by Beijing Academy of Artificial Intelligence by providing the computing resources.
• This project is advised by Cheng Ma.

Citation

@article{zhang2021deep,
  title={Deep-E: a fully-dense neural network for improving the elevation resolution in linear-array-based photoacoustic tomography},
  author={Zhang, Huijuan and Bo, Wei and Wang, Depeng and DiSpirito, Anthony and Huang, Chuqin and Nyayapathi, Nikhila and Zheng, Emily and Vu, Tri and Gong, Yiyang and Yao, Junjie and others},
  journal={IEEE Transactions on Medical Imaging},
  volume={41},
  number={5},
  pages={1279--1288},
  year={2021},
  publisher={IEEE}
}

@software{di_kong_2025_17364359,
  author       = {Di Kong},
  title        = {OpenDeepE: Open-Source Deep-E / 2D Fully-Dense U-Net},
  month        = oct,
  year         = 2025,
  publisher    = {Zenodo},
  version      = {v1.0},
  doi          = {10.5281/zenodo.17364359},
  url          = {https://doi.org/10.5281/zenodo.17364359}
}