Deep learning tomographic reconstruction for Photon Counting Computed Tomography (PC-CT)

Stage numéro : M2-1920-IM-01
Laboratoire :Centre de Physique des Particules de Marseille Case 902
 163 avenue de Luminy - 13288 Marseille Cedex 9
Directeur :Cristinel Diaconu - 04.91.82.72.01 - diaconu@cppm.in2p3.fr
Correspondant :William Gillard - 04.91.82.72.67 - gillard@cppm.in2p3.fr
Groupe d'accueil :imXgam
Chef de groupe :Christian Morel - 04.91.82.76.73 - morel@cppm.in2p3.fr
Responsable de stage :Yannick Boursier - 0033 4 91 82 76 41 - boursier@cppm.in2p3.fr

Thématique : Imagerie médicale

Master Internship / possible PhD position with an industrial partner

Keywords: Tomographic reconstruction, deep learning, optimization, Photon-Counting CT

Place and academic partner: Centre de Physique des Particules de Marseille, Aix-Marseille Univ., Marseille, France

Industrial partner: Technologies de France, Gardanne, France.

(Preferred) Starting date: February/March 2020 (up to 6 months). Available now.

The imXgam group of the Centre de Physique des Particules de Marseille (CPPM) has developed the micro PC-CT PIXSCAN-FLI, based on the hybrid pixel technology camera XPAD3 [1, 4] working in Photon Counting mode. The XPAD3 circuit permits to count single X-ray photons and to set a minimum threshold on their energy. A Monte Carlo simulation, which reproduces the XPAD3 response, has already been set up [2].

A broad variety of iterative algorithms have been developed for past years to perform 3D tomographic reconstruction of Photon Counting CT, among which those developed at CPPM for standard PC-CT [3]. They all rely on optimization techniques minimizing a cost function. This latter classically encompasses the physics of acquisition and a regularization term modeling the class of images to recover. However, the advent of deep learning approaches for CT reconstruction recently permitted a significant gain of quality when compared to the gold standards of iterative algorithms. Firstly used as denoisers, the deep networks very recently evolved so that their architecture mimics the one of iterative algorithms while offering better implicit modeling of images.

The subject of the internship is to develop and assess a 3D reconstruction method relying on deep learning techniques for Photon Counting CT. The implemented algorithm(s) will be based on state-of-the-art [5,6]. Their performance for Photon-Counting CT will be evaluated on simulated data (analytic or/and Monte-Carlo simulation) and on real data. The candidate will acquire real data on the micro PC-CT PIXSCAN-FLI at CPPM. This internship will be in collaboration with the company “Technologies de France” that developed its own spectral PC-CT prototype OSIRX, real data will be also acquired on ORISIX in order to validate the approach on several PC-CT.

Skills:

• Master degree in Applied Mathematics or Data science or Medical Imaging Science.

• Appreciated: Python or Matlab, knowledge of X-ray physics and related challenges.

Application: send to boursier@cppm.in2p3.fr AND y.menjour@compagnie-france.fr

• Curriculum Vitae

• Contact details for one or two referees

• Recent university records

References:

[1] Delpierre, P. et al., PIXSCAN : Pixel detector CT-scanner for small animal, Nucl. Instrum. and Meth. A 571 (2007) 425-428.

[2] F. Cassol Brunner et al., Simulation of PIXSCAN, a photon counting micro-CT for small animal imaging, JINST 4 (2009) P05012.

[3] Anthoine et al., Some proximal methods for Poisson intensity CBCT and PET, Inv. Prob. Imag., 4(4), (2012)

[4] F. Cassol et al., Characterization of the imaging performance of a micro-CT system based on the photon counting XPAD3/Si hybrid pixel detectors, Biomed. Phys. Eng. Express 2 (2016) 025003

[5] J. Adler, O. Öktem, Learned Primal-Dual Reconstruction, IEEE Trans. Med. Imag., 37 (6) (2018)

[6] Kyong Hwan Jin et al., Deep Convolutional Neural Network for Inverse Problems in Imaging, IEEE Trans. Imag. Proc. 26 (9) (2017)