Reconstructing volumetric Computed Tomography data from 2D X-ray projections with deep learning
Principal Investigator
Name
Tianyu Han
Degrees
M.Sc
Institution
Physics of Molecular Imaging Systems, Experimental Molecular Imaging, RWTH Aachen University
Position Title
Research assistant
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-717
Initial CDAS Request Approval
Oct 13, 2020
Title
Reconstructing volumetric Computed Tomography data from 2D X-ray projections with deep learning
Summary
Computed tomography (CT) is a powerful tool for diagnosing pathologies due to they are able to render volumetric information.
However, patients are always exposed to a higher dose of radiation in a CT machine and it's generally more costly than a common X-ray exam.
Given the fact human anatomy are well constrained, we are interested to extract the maximum amount of organ information from a 2D X-ray projection with the help of machine learning.
Efforts have been made to reconstruct 3D CT scans from single- and biplanar X-rays with machine learning approaches.
Nevertheless, the paired Xray-CT datasets of all established approaches are expensive to collect and always misaligned.
In our project, we will investigate the feasibility of transferring 2D X-ray data to 3D CT volumes without paired information.
However, patients are always exposed to a higher dose of radiation in a CT machine and it's generally more costly than a common X-ray exam.
Given the fact human anatomy are well constrained, we are interested to extract the maximum amount of organ information from a 2D X-ray projection with the help of machine learning.
Efforts have been made to reconstruct 3D CT scans from single- and biplanar X-rays with machine learning approaches.
Nevertheless, the paired Xray-CT datasets of all established approaches are expensive to collect and always misaligned.
In our project, we will investigate the feasibility of transferring 2D X-ray data to 3D CT volumes without paired information.
Aims
We would like to test the feasibility of transferring 2D X-ray data to 3D CT volumes without paired information.
Such a transition will be accomplished by training a deep neural network.
A large CT dataset, i.e., NLST, is a must to sufficiently train all parameters of our deep neural networks.
Collaborators
Dr. med. Dipl.-Phys. Daniel Truhn, UKA
Prof. Dr.-Ing. Volkmar Schulz, RWTH