DATE:

Friday, February 17, 2023

TIME:

3:30 PM

LOCATION:

GMCS 314

SPEAKER:

Dr. Hajar Homayouni, Computer Science, San Diego State University

ABSTRACT:

Chest X-ray images provide critical information for the diagnosis of COVID-19. Machine learning techniques for COVID-19 detection require substantial amounts of chest images to discover correct patterns. However, concerns over confidentiality and privacy have limited access to patients’ data. The distribution of samples across normal/abnormal classes is typically biased or skewed due to unavailability of sufficient data because of COVID-19 recency. Existing synthetic COVID-19 data generation approaches fail to generate high-resolution and diverse images. Moreover, there is a lack of research identifying whether synthetic images represent patients at high risk of severe disease, which is critical for making treatment decisions. We propose a High-Resolution COVID-19 X-Ray Generator (HRCX) framework based on a combination of a generative adversarial network and a predictive learning model that uses limited available chest images to generate balanced diverse high-resolution COVID-19 images with their severity scores. We use StyleGAN2 with adaptive discriminator augmentation, which controls generated images’ style and generates diverse patterns. In addition, we provide a COVID-19 severity index to aid in predicting illness severity. We generated 3300 high-quality and diverse COVID-19 X-Ray images with a resolution of 512×512, which we further increased to 1024×1024 with the help of Super-Resolution. Additionally, severity scores of 300 images are calculated and demonstrated to be effective in both normal and infected cases.

HOST:

Bryan Donyanavard

VIDEO: