Transfer-Learning-Stenosis-Detection

---

In this repository you can find the code implementation for the paper "Transfer Learning for Stenosis Detection in X-ray Coronary Angiography", published on Mathematics 2020, Volume 8, Issue 9, 1510: https://www.mdpi.com/817954

This paper presents a new method for the automatic detection of coronary artery stenosis in XCA images, employing a pre-trained Convolutional Neural Network (CNN) via Transfer Learning. The method is based on a network-cut and fine-tuning approach.

---

Release notes

All the experiments were computed on a Cloud-Platform with an Intel (R) Xeon (R), 12 GB of RAM, and a 2.00 GHz dual processor, and a Tesla P4 GPU with 8 GB VRAM. The algorithms were implemented in Pytorch 1.8 library and Python 3.7.10.

The source code is in constant update. Stay close to updates.

• 2021.04: Release release/1.8-pytorch version

Aditional packages: torchsummary and livelossplot.

---

Training stage

The training XCA images were patches of pixels, such that

• stenosis_data
  • train
    • positive
    • negative
  • validation
    • positive
    • negative

Supported architectures (so far):

• ResNet (resnet18, resnet34, resnet50)
• VGG (vgg11, vgg13, vgg16)

ResNet's Architectures and configurations:

(0a):  conv1
(0b):  bn1
(0c):  relu
(0d):  maxpool
(1):  layer1
(2):  layer2
(3):  layer3
(4):  layer4
(pooling):  avgpool
(fc):  fc1

Cut-blocks parameters, e.g., 1 means that the network will be cut in the output of the layer1 (fisrt residual block), and then connected to the avgpool.:

• [1, 2, 3, 4]

Fine-tuning layers parameters, from top to bottom, e.g., 1 means only the fc1 layer will be finetuned.

Vgg's Architectures and configurations:

(0):  conv1
(1):  conv2
(2):  conv3
(3):  conv4
(4):  conv5
(pooling):  avgpool
(fc):  fc

Cut-blocks parameters, e.g., 1 means that the network will be cut in the output of the conv2 (first convolutional block), and then connected to the avgpool.:

• [1, 2, 3, 4]

Fine-tuning layers parameters, from top to bottom, e.g., 1 means only the fc1 layer will be finetuned, and 7 that all the layers will be finetuned:

For all models, if a -1 is passed as finetuning layer, all the layers will be finetuned.

All the configuration parameters are in a text file, such as

[PARAMS]
DATA_DIR = 'DATASET FOLDER'
WEIGHTS_DIR = 'DIRECTORY TO SAVE WEIGHTS '
torch_seed = 'PYTORCH SEED'
model_type = 'MODEL TYPE: resnet/vgg'
model_deep = 'MODEL DEPTH'
model_name = 'MODEL FILENAME'
cut_block = 'CUT BLOCK: 1-4'
train_layers = 'NO. TRAINABLE LAYERS'
batch_size = 'TRAINING BATCH SIZE'
pretrained = 'USE IMAGENET: True/False'
lr = 'LEARNING RATE'
momentum = 'MOMENTUM
factor= 'LEARNING RATE DECAY FACTOR'
patience = 'VAL_LOSS PATIENCE'
num_epochs = 'NUM OF EPOCHS TO TRAIN'
finetuning = 'FINETUNE A PREVIOS MODEL, READED FROM WEIGHTS_DIR'

Running example (in a coolab jupyter notebook):

%run main_train.py trian_params.txt

It generates in the WEIGHTS_DIR folder:

• A .pth file with the weights of the model
• A .json file with the training acc/loss history

---

Testing stage

If you want to test in your own dataset, set on a train folder all the XCA images. Notice that the size must be pixels, and separate into two subfolders: positive and negatives cases. For instance:

• stenosis_data
  • train
    • positive
    • negative

Weights example:

• model_resnet18_3RB_ALL_FT.pth : resnet18 weights with three residuals blocks (of four) with all the layers finetuned.

All the configuration parameters are in a text file, such as

[PARAMS]
DATA_DIR = 'DATASET PATH'
WEIGHTS_DIR = 'DIRECTORY TO SAVE WEIGHTS '
model_type = 'MODEL TYPE: resnet/vgg'
model_deep = 'MODEL DEPTH'
model_name = 'MODEL FILENAME'
cut_block = 'CUT BLOCK: 1-4'
train_layers = -1
RESULTS_DIR = 'RESULTS PATH'
GRADCAM = 'APPLY GRADCAM: True/False'
imagenet_norm = 'APPLY IMAGENET NORMALIZATION: True/False'
cam_layer = 'GRADCAM LAYER'

Running example (in a coolab jupyter notebook):

%run main_test.py test_params.txt

It generates in the WEIGHTS_DIR folder (where the weights are loaded)

• A .json file with the testing report from each metric
• A .csv file with the testing filename of the image, the ground-truth label, the predicted probability, and the predicted label, respectively.

Detection results may change from the previously reported results due to the random seed, random split test/validation/train dataset:

---

Cite as

If you use this for research, please cite. Here is an example BibTeX entry:

@article{ovalle2020transfer,
  title={Transfer Learning for Stenosis Detection in X-ray Coronary Angiography},
  author={Ovalle-Magallanes, Emmanuel and Avina-Cervantes, Juan Gabriel and Cruz-Aceves, Ivan and Ruiz-Pinales, Jose},
  journal={Mathematics},
  volume={8},
  number={9},
  pages={1510},
  year={2020},
  publisher={Multidisciplinary Digital Publishing Institute}
}

---

Development

Want to contribute? Great!. Contact us.

Follow our research work at: https://emmanuelovalle.netlify.app/

---

License

MIT