CCS-GAN: COVID-19 CT Scan Generation and Classification with Very Few Positive Training Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Training deep learning models for COVID-19 diagnosis is challenging due to the scarcity of positive CT scan images.
• Existing methods require large datasets, hindering the development of effective diagnostic tools, especially during pandemics.
• Limited positive training data impedes the application of deep learning in medical image analysis for rare or emerging diseases.

Purpose of the Study:

• To develop a novel algorithm capable of generating high-quality synthetic COVID-19 pneumonia CT scan slices.
• To enable accurate classification of COVID-19 pneumonia using deep learning models trained on a minimal number of positive training images.
• To reduce the data acquisition burden and facilitate data sharing between healthcare institutions.

Main Methods:

• Introduction of the cycle-consistent segmentation-generative adversarial network (CCS-GAN).
• Integration of style transfer, pulmonary segmentation, and transfer learning from normal CT scans.
• Generation of synthetic positive COVID-19 CT images using a small dataset of actual positive and a larger dataset of normal images.

Main Results:

• The CCS-GAN algorithm successfully generated synthetic COVID-19 CT scan slices of sufficient accuracy.
• A deep neural network (DNN) classifier, augmented with CCS-GAN synthetic data, achieved high classification accuracy.
• The system demonstrated high performance using as few as 10 positive training slices, a significant reduction compared to existing methods.

Conclusions:

• The proposed CCS-GAN algorithm effectively addresses the challenge of limited positive training data in medical deep learning.
• This approach significantly lowers the barrier for training accurate COVID-19 diagnostic classifiers, improving accessibility to AI-driven diagnostics.
• The method holds promise for advancing research in medical image analysis and disease screening, particularly for conditions with data scarcity.