3D GANs for creating synthetic medical volumes
Title: 3D GANs for creating synthetic medical volumes
SNIC Project: Berzelius-2021-93
Project Type: LiU Berzelius
Principal Investigator: Anders Eklund <anders.eklund@liu.se>
Affiliation: Linköpings universitet
Duration: 2021-12-14 – 2022-07-01
Classification: 20603
Homepage: https://liu.se/forskning/assist
Keywords:

Abstract

Deep learning is currently revolutionizing many research fields. In computer vision, considerable progress has been made during the last 5 years, and a crucial resource is the ImageNet database (Deng et al., 2009) which contains more than 14 million labeled images. Any researcher can use ImageNet as training data to improve methods in deep learning based computer vision. Techniques developed in computer vision are rapidly transferred to the medical imaging field, but a major constraint is that access to medical images is more complicated due to ethics and data protection legislation (e.g. the General Data Protection Regulation (GDPR)). There are a number of openly available medical imaging datasets, but they are much smaller compared to ImageNet. Health care providers have records containing vast quantities of medical images, but these records are often not accessible for research due to regulatory hurdles. In this project, we will therefore develop 3D generative adversarial networks (GANs) to synthesize realistic medical volumes. As synthetic volumes are not attributable to a specific person, our hypothesis is that data protection legislation does not apply and the data can therefore be shared freely. Our vision is a VolumeNet database, which will contain millions of realistic synthetic medical volumes reflecting true diversity of scans. The database will accelerate deep learning research in medical imaging, similarly to how ImageNet has accelerated deep learning research in computer vision. Goal: Synthesize realistic medical volumes (e.g. brains) using 3D GANs. The generated volumes can for example be used for training a classifier (e.g., classify a brain as healthy or diseased) or for training image segmentation algorithms. For image segmentation, realistic volumes as well as the corresponding ground truth segmentations / annotations will be synthesized. New metrics for evaluating how good a 3D GAN is will be developed.