PhD Fellowship: Deep learning-based algorithms for spectral CT
Recrutement: 
Recrutement en cours/passé: 
Recrutement passé
Periode: 
2018-2021
Contact: 
francoise.peyrin@creatis.insa-lyon.fr, nicolas.ducros@creatis.insa-lyon.fr, Abascal juan.abascal@creatis.insa-lyon.fr

Keywords   Deep learning, convolutional neural network, spectral computed tomography (CT).

Context   CREATIS is a research unit of CNRS/INSERM/INSA Lyon/University of Lyon devoted to medical imaging. The candidate will join the Tomographic Imaging and Radiotherapy team, which has internationally recognized expertise in X-ray imaging and inverse problems.

Project    Spectral computed tomography (CT) is a new imaging modality that can resolve the concentration of the constituents of the human body (e.g., bone, water, fat) or contrast agents [1].  The spectral CT reconstruction problem is usually addressed as a (nonlinear) inverse problem, which requires the knowledge of source and detector response functions [2]. However, these are generally unknown or difficult to model.

We propose to overcome these difficulties by constructing new reconstruction algorithms based on deep learning. Deep learning has been forecasted as one the 10 breakthrough technologies of 2017 [3] and is proving to be one of the most powerful techniques in computer vision, with promising results in biomedical applications [4]. Just recently, several authors proposed to use these techniques for learning inverse problems [5], [6].

Research Program   The goal of this thesis is to develop new algorithms based on deep learning for improving image quality in spectral CT. There are two specific objectives: learning the nonlinearities and circumvent modelling the source and detector response functions, and designing specific deep iterative learning algorithms for spectral CT. We will investigate various deep learning architectures [3] and compare them to model-based approaches. The successful candidate will:

  • Contribute to the development of our in-house Matlab toolbox
  • Implement deep learning methods using TensorFlow
  • Compare with existing methods using simulated and experimental data
  • Collaborate with CPPM (Marseille) and UCL (CMIC, London, UK)

Skills        The student must have a strong background image processing and deep learning. Knowledge in inverse problems, medical imaging and radiation physics is a plus. Programming skills: Matlab, Python.

Practical information

  • The thesis will take place at CREATIS, Lyon, France.
  • Three-year funding starting in September 2018.
  • The salary about €1500 net monthly (doctoral contract from the doctoral school EEA de l’Univerité de Lyon)

How to apply?  

Send both your CV and academic records to

Reference

[1] K. Taguchi et al, “Vision 20/20: Single photon counting x-ray detectors in medical imaging,” Medical Physics, 40, 100901, 2013.

[2] N. Ducros et al. “Regularization of nonlinear decomposition of spectral X-ray projection images” Medical Physics, 44, 9, e174-e187, 2017.

[3] “10 Breakthrough Technologies 2017 - MIT Technology Review.” [Online]. Available: https://www.technologyreview.com/lists/technologies/2017/.

[4] Y. LeCun et al. “Deep learning”, Nature 521, 436–444, 2015.

[5] O. Öktem and J. Adler, “Solving ill-posed inverse problems using iterative deep neural networks,” Inverse Probl., 2017.

[6] S. Yu et al., “Deep De-Aliasing for Fast Compressive Sensing MRI,” May 2017.