Classic Augmentation Based Classifier Generative Adversarial Network (CACGAN) for COVID-19 Diagnosis

• Trained the Generator with a custom loss function to enable it to generate new data in specific classes
• Trained the Discriminator with the original data set and the data generated by the Generator
• Evaluated the performance of the GAN model by multiple classifiers (VGG, ResNet, EfficientNet, etc.)

Dataset Links

• covid19-image-dataset

Configuration

• conda env create -f environment_dl.yml
• environment_dl.yml may contains some packages that won't be used here! If you are disk-sensitive, please only install the packages appeared in scripts :)

Usage

• ClassicHistEqual.ipynb for histogram equalization and ClassicAUG.ipynb for classic augmentation
• CACGAN_AUG.ipynb for data synthetizing using an AC-GAN model
• GenerateData.ipynb for generating new data to ./GANGEN through the Generator saved during training
• folders [augmented, original, synthetic] are results of multiple classifiers on augmented data, original data, and synthetic data, respectively

Classic Augmentation

Histogram Equalization

Lung CT Images Generated by the Generator After 1, 20, 50, 100, 500, 1000 Epochs

Generated Data of the Generator After the Last Iteration (Left), and the Real Data (Right)

Structure of Generator and Discriminator

Training Process