/PAM_de-noising

Primary LanguagePython

PAM de-noising tool

Da He, Jiasheng Zhou, Xiaoyu Shang, Xingye Tang, Jiajia Luo, Sung-Liang Chen

De-Noising of Photoacoustic Microscopy Images by Attentive Generative Adversarial Network (Paper link)

Citation:

If you find this implementation, the inference tool, or the article is helpful / useful / inspiring, please cite the following :D

@article{he2022noising,
title={De-Noising of Photoacoustic Microscopy Images by Attentive Generative Adversarial Network},
author={He, Da and Zhou, Jiasheng and Shang, Xiaoyu and Tang, Xingye and Luo, Jiajia and Chen, Sung-Liang},
journal={IEEE Transactions on Medical Imaging},
year={2022},
publisher={IEEE}
}

This implementation was developed based on the repository https://github.com/eriklindernoren/Keras-GAN/tree/master/srgan.

Dependencies

  • tensorflow
  • keras
  • numpy
  • keras_contrib
  • scipy==1.2
  • Only GPU-based calculation is supported now.

Inference Utilization

  • Step 1: Place all the noisy image (in grayscale .png format only) in a folder ($input_dir$)

  • Step 2: Under the root folder of the tool, input the command:

python inference.py --input_dir $input_dir$

  • Step 3: The de-noised results will be place in $input_dir$/denoised_out

  • Note: Image shapes in the input folder can be arbitrary and different.

Training Utilization

  • Step 1: Prepare the training set and validation set in .npy format with 0~1.0 value range. For each set, noisy input data and clean groundtruth data should be placed in noisy and clean sub-folders individually with the same file names.

  • Step 2: Modify Lines #16, #23, #55, #62 in the "data_loader.py" file to specify the dataset.

  • Step 3: Hyper-parameters of the training could be modified in Lines #132, #142, #147, #514 in the "pam_denoise_main.py" file.

  • Step 4: Run the command python pam_denoise_main.py to start training.

  • Step 5: Training results will be saved in "./saved_model".