endscope_diagnosis

learning

1. 10 text files should be located as follows. Each text file corresponds to each sample of learning data.

text
└── 3N100
    ├── 3N100_01.txt
    ├── 3N100_02.txt
    ├── 3N100_03.txt
    ├── 3N100_04.txt
    ├── 3N100_05.txt
    ├── 3N100_06.txt
    ├── 3N100_07.txt
    ├── 3N100_08.txt
    ├── 3N100_09.txt    
    └── 3N100_10.txt

The format of 3N100_{num}.txt is defined as follows:

{image_path} {label(hyperplasticpolyp=0,Adenoma=1,Normalmucosa=2)}
{image_path} {label(hyperplasticpolyp=0,Adenoma=1,Normalmucosa=2)}
{image_path} {label(hyperplasticpolyp=0,Adenoma=1,Normalmucosa=2)}
:
:

example:

../../Images/cropped/A16/A16-00097.jpg 0
../../Images/cropped/A16/A16-00098.jpg 0
../../Images/cropped/A16/A16-00099.jpg 0

2. learning

You can use monitor.py for training neural networks. exptmgr.py is a script for the 10-fold cross-validation, namely, a single run of exptmgr.py can generate 10 models in turn.

example:

cd ResNet_tf2/sorce
python monitor.py ../text/3N100/3N100_01.txt ./config.ini

cd ResNet_tf2/sorce
python exptmgr.py ../text/3N100 ./config.ini -c

validation

inspector.py yields the list of the label of teacher value and corresponding output by using the trained model for misclassified test data.

cd ResNet_tf2/sorce
python inspector.py ../products/product_3N100/wh5/3N100_01_w.h5 ../text/3N100/3N100_01.txt

gradcam.py is for GradCAM.

cd ResNet_tf2/sorce
python gradcam.py ../products/product_3N100/json/3N100_01.json ../products/product_3N100/wh5/3N100_01_w.h5 ../../Images/cropped/A01/

igen.py, iread.py, and iseq.py are used for preprocessing of input images.

compatibility

• Ubuntu 18.04.5 LTS
• CUDA 10.0
• CUDNN 7.6.5.32
• Python 3.6.9
• numpy 1.20.0.dev0+8b15e57
• opencv-python 4.1.0.25
• tensorflow-gpu 2.0.0