README

Code for the paper: "Deep-learning-aided forward optical coherence tomography endoscope for percutaneous nephrostomy guidance"[1] The following pieces of python code and jupyter notebooks were used for the paper. The following architectures were used:

• Resnet 34
• Resnet 50 and Mobilenetv2 with and without pretrained initial weights from Imagenet Dataset.

The dataset can be found in [2]

Prerequisites

The language used is Python. We used Tensorflow 2.3.

Structure:

• 0-Read_images.ipynb
  It process the images from JPEG to numpy ndarray binaries

• ResNet34/
  

  • Cross-validation
    • archResNet_p1.py
      
    • archResNet_p2.py
      
    • archResNet_p3.py
      
    • archResNet_p4.py
      

  It uses the ResNet34 architecture to predict the type of tissue( 3 categories) It is split in 4 files in order to be able to run them independently.

• PT_MobileNetv2/
  

  • Cross-validation
    • PT_MobileNetv2_batch/
      
      • mobilenetv2_tl_arg_simult_vC.batch
    • PT_MobileNetv2_python/
      
      • mobilenetv2_tl_arg_vC.py

• ResNet50/
  

  • Cross-validation
    • Resnet50_batch/
      
      • resnet50_arg_simult.batch
        
    • Resnet50_python/
      
      • archResNet50_arg.py
        
  • Cross-testing
    • Resnet50_batch/
      
      • resnet50_arg_outer_simult.batch
        
    • Resnet50_python/
      
      • archResNet50_arg_outer.py
        

• PT_ResNet50/
  

  • Cross-validation
    • PT_Resnet50_batch/
      
      • resnet50_tl_arg_simult.batch
        
    • PT_Resnet50_python/
      
      • archResNet50_tl_arg.py
        
  • Cross-testing
    • PT_Resnet50_batch/
      
      • resnet50_tl_arg_outer_simult.batch
        
    • PT_Resnet50_python/
      
      • archResNet50_tl_arg_outer.py
        

• Processing_results.ipynb
  Processing of the results to obtain the accuracies, epochs of all the combinations. Time is calculated for a few combinations

• Processing_predictions.ipynb
  Processing of the predictions to obtain the ROC curves

• Processing_time.ipynb
  Complete processinf of time for cross-validation.

• Grad-CAM.ipynb
  Implementation of visual explanation using Grad-CAM[3] for the models obtained

For ResNet34 run the python code, for the rest you need to use arguments. The python file is used as:

archResNet50_arg.py testing_kidney validation_kidney

e.g.

archResNet50_arg.py 1 2

The batch file was used in Summit supercomputer.

Paper

[1] Chen Wang, Paul Calle, Nu Bao Tran Ton, Zuyuan Zhang, Feng Yan, Anthony M. Donaldson, Nathan A. Bradley, Zhongxin Yu, Kar-ming Fung, Chongle Pan, and Qinggong Tang, "Deep-learning-aided forward optical coherence tomography endoscope for percutaneous nephrostomy guidance," Biomed. Opt. Express 12, 2404-2418 (2021)

Paper link

Dataset

[2] Chen Wang, Paul Calle, Qinggong Tang, & Chongle Pan. (2022). OCT porcine kidney dataset for percutaneous nephrostomy guidance [Data set]. Zenodo. https://doi.org/10.5281/zenodo.7113948

References

[3] Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-CAM: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (pp. 618-626).

Contact

Paul Calle - pcallec@ou.edu
Project link: https://github.com/thepanlab/FOCT_kidney