CACGAN

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

© Haodong Li

Overview

Lung images generated by the Generator after 1, 20, 50, 100, 500, 1000 epochs
  • Trained the Generator with a custom loss function to enable it to generate new data in specific classes, and the Discriminator with the original data set and the data generated by the Generator.
  • Evaluated the performance of CACGAN by multiple classifiers (VGG, ResNet, EfficientNet, etc.)

Motivation

  • It is difficult to easily obtain large quantities of well-labelled data from hospitals.
  • In order to optimize the generalization ability while reducing the computational costs, this paper proposes a CACGAN for data synthetizing.

Dataset

Details of the dataset
Some lung images in the dataset

Methodology

Classic augmentation (left) and histogram equalization (right)
  • Adopt classic augmentation to avoid overfitting,
  • and histogram equalization to enhance the feature through improving the contrast of the whole image.
Structure of Generator and Discriminator
Training process

Experiments

Generated data of the Generator after the last iteration (left), the real data of corresponding batch (mid), and the training log (right)
Performance of classifiers on original data
Performance of classifiers on classic augmented data
Performance of classifiers on generator synthetic data
  • Although the performance of generated data is slightly inferior, it can still be seen that the CACGAN can generate lung images very effectively.
  • This may be an effective method for solving the problem that it is usually difficult to obtain large quantities of well-labeled data in the medical field.