About

The COVID-19 pandemic has caused major disruptions to life around the world. It has become critical for healthcare workers to be able to quickly and accurately test patients for COVID-19 so that patients with the disease can be quickly quarantined and treated. While many different methods of performing COVID-19 tests exist, one method is to use computed tomography (CT) scans of patients to determine whether they have COVID-19. Such evaluation can be expedited through the use of computer aided diagnosis (CAD).

Deep learning models provide a method of performing CAD. By training a model to classify the CT scans of patients as positive or negative for COVID-19, patients who have CT scans performed, regardless of the original intent of the screening, can be easily tested for COVID-19. However, training such models requires a large enough dataset so that the model can be sufficiently trained without overfitting.

This project seeks to address that issue by developing a synthetic dataset of CT scans for patients with COVID-19. These synthetic CT scans would be available to the public (on this website) so that other researchers may use them in developing their own models. A synthetic dataset would be advantageous in that it can be neatly organized for easy use in training a model and in that it would not correspond to actual patient medical records, avoiding any issues of privacy.

We look to create this synthetic dataset using a generative adversarial network (GAN). By using real datasets, both public and private, of CT scans of patients with COVID-19, we can train a GAN to create realistic synthetic CT scans.

Early Results

Below are some of our preliminary results. We will update this page with more results as the project progresses. More recent results are listed first.

These images are our recent results for generating axial slices from CT scans for patients with COVID-19. We use graph-based segmentation to generate only the lungs, then we paste those lungs into a real image. Our conditional GAN model to generate the lungs uses the mean teacher algorithm and a modified relativistic loss function.

These are our most recent results for generating COVID-19 X-ray images. To generate these images we used a mean-teacher algorithm, transfer learning, and data augmentation. Transfer learning was applied using a dataset of normal chest X-rays and X-rays for patients with pneumonia. Due to memory constraints, these images are of resolution 128x128.

These images are early results for generating X-ray images of patients from a pneumonia dataset. These images were generated using a mean-teacher algorithm. These images are an intermediate result for using tranfer learning to generate X-ray images for patients with COVID.

These are some early results for generated synthetic X-ray imagery. These images were created using a GAN after applying transfer learning using a non-COVID dataset. We are currently working on generated higher-quality and higher-resolution images.