Advanced visual turing test of Generative Adversarial Networks using lung CT scans
Principal Investigator
Name
Pascal Michel
Degrees
M.Sc.
Institution
Michel Ventures UG (haftungsbeschränkt)
Position Title
CEO & Research Lead
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-535
Initial CDAS Request Approval
Jul 11, 2019
Title
Advanced visual turing test of Generative Adversarial Networks using lung CT scans
Summary
Generative Adversarial Networks (GANs) have been used to synthesize medical images with moderate to good authenticity. This project aims to study the distinguishability of nuanced medical images (lung CTs) and their synthetic counterparts in a saturated setting (using a large data set). Therefor, we will train Deep Convolutional Generative Adversarial Networks (DCGANs) on the NLST data set and subsequently generate a purely synthetic image data set (2D, 512 x 512). This "fake" data set will be mixed with "real" 2D slices from the NLST data set and given to a senior radiologist in a random sequence with the task to identify images as "real" or "fake" (visual Turing test).
The hypothesis we aim to validate is, that an experienced radiologist is not able to distinguish between synthetic and real data, indicating that GANs preserve highly nuanced clinical information in artificial data sets.
The hypothesis we aim to validate is, that an experienced radiologist is not able to distinguish between synthetic and real data, indicating that GANs preserve highly nuanced clinical information in artificial data sets.
Aims
- Train a DCGAN on the NLST and succeed in a visual Turing test on lunc CT data for the first time
- Qualitatively demonstrate the capacity of GANs to generate artificial data sets that preserve highly nuanced clinical features
Collaborators
PD Dr. med. Tobias Penzkofer (Charité Berlin)
Prof. Dr. rer. Felix Biessmann (Beuth University of Applied Science)