COSMIC Compressed Sensing for Magnetic Resonance Imaging Cosmology

COSMIC: Compressed Sensing for Magnetic Resonance Imaging & Cosmology Philippe Ciuciu (I 2 BM) & Jean-Luc Starck (IRFU) October 19 th, 2016 UNIRS | 03 -02 -2016 | PAGE 1

MRI CONTEXT IN COSMIC

RADIO-INTERFEROMETRY CONTEXT IN COSMIC

COMMONALITIES BETWEEN MRI & RADIOINTERFEROMETRY • Data sampling in the 2 D/3 D Fourier domain: k-space MR Image Linear Reconstruction Acquisition IFFT 7 Tesla MRI Scanner @Neuro. Spin Spatial frequencies

COMMONALITIES BETWEEN MRI & RADIOINTERFEROMETRY • Sampling in the 2 D Fourier domain:

COMMON ISSUE: NEED FOR SPARSE SAMPLING • High resolution imaging requires collecting too much data • Temporally-resolved hyperspectral imaging necessary to recover fast transient sources A common need for Compressed Sensing Data is sparse, compressible, redundant… Sense the compressed information directly! Donoho, Tao, Romberg, Candes

COSMIC: SCIENTIFIC OBJECTIVES Ø Magnetic Resonance Imaging ime Compressed Sensing (CS-MRI) t n st itio m co. MR s i Fourier space u q e exa ld Image c a ce ecreas etic fie u d e ort, d magn R : I , … igh cts e MR nt comf Nonlinear h m t i n t i e s a artifa e, kes pati sue t pac Data 7 Tesla (Neuro. Spin) acquisition 7 T MRI scanner@Neuro. Spin Sta prove ting tis vemen ion in s Im ea lut mo h o void atient’s igh res. Image A h reconstruction p • imit e very L v • chie A • Random undersampling Sparsity in wavelet basis • Ø Radio-interferometry Fourier space omy Galaxy time image (Cygnus A) ce, a p ron tion in s t s A lu ube c adio igh reso Nonlinear l of R a r h t n c c i h o s Data es e very yperspe source the Ep k a t S chiev uct h sient ve to i r • A econst of tran al relat n n Image acquisition • R etectio the sig D e t reconstruction a • stim ation E • nis reio

COSMIC: MAJOR STAKES FOR CEA & DRF • MRI (11. 7 Tesla magnet at Neuro. Spin, ISEULT project) Ø 60 millions of investment for the strongest machine worldwide Ø Optimal exploitation based on CS: imaging better and faster Ø Disruptive technology transferable to industry (Siemens, GE …) Ø Partnerships with Europe (Germany) and US (CMRR) • Radio-interferometry (Squared Kilometer Array telescope) Ø Positioning on the biggest array telescope worldwide Ø Big data analysis (Data deluge: SKA will acquire 9 Tb/s, more than Google !!! and 1 PB per day in the archive)

COSMIC: ISSUES & LOCKS UP (1/4) • MRI: Optimize image representation & increase sparsity Curvelets Starlets, shearlets Cosmo. Stat Neuro. Spin

COSMIC: ISSUES & LOCKS UP (2/4) • MRI: Optimize image reconstruction • Synthesis formulation (eg, FISTA): • Analysis formulation (eg, ADMM): Observation model: • Equivalence between synthesis & analysis formulations for wavelet bases • Faster algorithms for the analysis formulation with redundant transforms • Extensions to 3 D imaging & multi-channel data Cosmo. Stat Neuro. Spin

COSMIC: ISSUES & LOCKS UP (3/4) • Radio-interferometry: Optimize measurement matrix Neuro. Spin Cosmo. Stat

COSMIC: ISSUES & LOCKS UP (4/4) • Computational aspects: Ø Scalability of 2 D reconstruction algorithms to 3 D imaging Ø Addressing the case of non-Cartesian Fourier data Ø Parallel programming on multi-CPU and GPU architectures using the most convenient languages (Python, Julia) on top of C/C++ libraries Neuro. Spin Cosmo. Stat

COSMIC: AGENDA & DELIVERABLES Hiring master trainees: - Sylvain Lannuzel (ECP+M 2 R) - … Cosmo. Stat Neuro. Spin D 1: Improved 2 D reconstruction 09/2016 D 4: (2 D+t) & 3 D imaging D 3: In-vivo Human acquisitions 12/2016 04/2017 09/2017 10/2016 M 1: Kick-off Meeting M 2: Advances on MRI reconstruction … … M 3: Gain in performance & scalability D 5: Gadgetron demonstrator New Ph. D student: Loubna El Gueddari Neuro. Spin M 4: Wrap-up Cosmo. Stat D 2: Parallel programming (code analysis, acceleration)

COSMIC CONSORTIUM: NEUROSPIN TEAM Neuro. Spin (UNATI-UNIRS) Ph. Ciuciu (leader) A. Vignaud Expertise in: - MR Physics & signal processing - Design of CS schemes - Non-Cartesian reconstruction - Parallelization of algorithms (GPU) C. Lerman CS in MRI Neuroimaging scientist C. Lazarus MR Physicists A. Coste Ph. D Student CS-MRI A. Grigis Engineer

COSMIC CONSORTIUM: COSMOSTAT TEAM Expertise in: - Astronomy and Signal processing - 2 D/3 D sparse decompositions - Modeling of prior informations - Inverse problem solving Astrophysics Division (Cosmo. Stat) J. -L. Starck (leader) J. Bobin F. Sureau CS in Astronomy Research Director S. Pires Sparse image analysis M. Jiang J. Girard Ph. D Student Euclid scientist CS in Astronomy Radio-astronomer

COSMIC CONSORTIUM: JOINING OUR EFFORTS Nurtering interactions in Compressed Sensing Neuro. Spin (UNATI-UNIRS) Ph. Ciuciu (leader) A. Vignaud Neuroimaging scientist C. Lazarus C. Lerman MR Physicists A. Coste Ph. D Student CS-MRI A. Grigis Astrophysics Division (Cosmo. Stat) J. -L. Starck (leader) J. Bobin Research Director S. Pires F. Sureau Sparse image analysis M. Jiang J. Girard Ph. D Student Engineer Euclid scientist CS in Astronomy Radio-astronomer

CONCLUSION