Skip to Main Content

An official website of the United States government

Using generative adversarial neural networks to create synthetic images for improved classification of lung cancer

Principal Investigator
Name
Caleb Bradberry
Degrees
Ph.D.
Institution
Radford University
Position Title
Assistant Professor
Email
About this CDAS Project
Study
NLST (Learn more about this study)
Project ID
NLST-392
Initial CDAS Request Approval
Feb 23, 2018
Title
Using generative adversarial neural networks to create synthetic images for improved classification of lung cancer
Summary
This project will use a neural network technique to generate synthetic lung cancer images and their surrounding data in an effort to enhance our understanding of machine learning classification for cancer. Specifically, this project applies Goodfellow's approach to neural networks, the generative adversarial neural network (GANN) (Goodfellow, 2014). GANN's provide a mechanism to generate synthetic images from real observation. By utilizing the images in the NLST, a GANN will be created that can generate synthetic images and their associated data, with the goal being to use these synthetic images to further enhance the abilities of all classes of neural networks to detect phenomena that could lead to cancer detection.
Aims

This project has three goals: to develop a neural network that can generate synthetic images, to use the synthetic images for replication and retraining models, and to test the retrained models on their ability to detect.

Goal 1: Train a generative adversarial neural network with the NLST to generate synthetic images. The outcome of this goal is the trained network itself that can then generate new images.
Goal 2: Using the images from goal 1, apply established techniques from research on the NLST itself to replicate their efficacy. Additionally, new models will be introduced that have the potential to lead to better classifiers.
Goal 3: Test the retrained models on their ability to detect. Additionally, test the models on their ability to discriminate real images from synthetic.

Collaborators

Wendy Downy - Radford University