"# covid-19-detector"

ATTENTION

Do NOT use any result of this repo for scientific purposes without medical assistance

Pre-requisites

We assume you have an GPU available and python 3.6+ installed. Make sure you have installed all the correct drivers for gpu training. Check out https://www.tensorflow.org/install/gpu for information about GPU setup.

Assuming you're ok, run pip install -r requirements.txt to install all necessary packages.

Training

Set the following options to train:

• -p: Path to the dataset (dataset/)
• -g: If use or not gpu in case it is possible. (Default = True)
• --network: Base network to use. Supports vgg16, vgg19, resnet50 and resnet152 (Default = resnet50)
• --hf: Augment with horizontal flips in training. (Default=True)
• --vf: Augment with vertical flips in training. (Default=True)
• --rot: Augment with 90 degree rotations in training. (Default=True)
• --bn: If use or not batch normalization. Available only for both vgg models
• --num_epochs: Number of epochs. (Default = 100)
• --config_filename: Name of txt file that stores all the metadata related to the training (to be used when testing)
• --input_weight_path: Input path for weights for classifier model
• --mn: Name of the model

Example of training command line:

python train_cnn.py -p dataset/ -g True --network vgg19 --mn vgg19_1

Predicting

Set the following options to predict:

• -w: Path to the weights file (hdf5 or h5)
• -c: Path to the config file
• -p: Path to the folder containing images to be classified
• -g: Use GPU or not (Default = True)

Pattern Visualization

Set the following options to run the vis script:

• -p: Path to the image
• -w: Path to the weights file (hdf5 or h5)
• -c: Path to the config file
• -g: Use GPU or not (Default = True)

Grad-CAM

Occlusion Sensitivity

Server

To run the server just enter the following line on cmd:

python server.py -m path/to/model/weights/file

The weights file were generated after training

After, to make an http request run request.py:

python request.py -p path/to/folder/containing/images/to/be/classified

Results

All the models were run on a GPU RTX 2080 6GB. Each epoch took ~22s

• VGG19: ~96% accuracy and ~90% val accuracy after 29 epochs

Credits

• Many thank to Società Italiana di Radiologia Medica e Interventistica for providing many images
• Many thanks to https://www.medicalimages.com for providing some CT images of normal lungs
• Many thanks to Joseph Paul Cohen who makes part of the dataset available
• Many thanks to Adrian Rosebrock for providing some interesting code