The SIMRI project A versatile and interactive MRI simulator *

Transcription

1 COST B21 Meeting, Lodz, 6-9 Oct The SIMRI project A versatile and interactive MRI simulator * H. Benoit-Cattin 1, G. Collewet 2, B. Belaroussi 1, H. Saint-Jalmes 3, C. Odet 1 1 CREATIS, UMR CNRS #5515, U 630 Inserm, INSA Lyon, UCB Lyon 2 CEMAGREF, Food process Engineering Research Unit, Rennes 3 LMRMN-MIB, UMR CNRS 5012, UCB Lyon * JMR, 173 (2005) Context 2. SIMRI overview 3. Simulation results 4. Simulator implementation 5. Perspectives 2/32

2 1. Context MR Imaging Recent and complex technique Based on protons magnetization High static field + RF excitations Contrast imaging adapted to soft tissue Images +- artéfacts Objects (Chemical shift, susceptibility, motion) Imaging device (Field, RF inhomogeneity, gradient non linearity) 3/32 MRI simulation Better understanding of MRI images Pedagogic purposes Conception, calibration and test of MRI sequences MRI Images with a «ground truth» Artifacts impact and correction Image processing assessment (segmentation, quantification) Main works 1D MRI simulation [Bittoun-81] 2D MRI simulation [Olsson-95] Simulation with a distributed implementation [Brenner-97] 3D brain MRI simulation [Kwan-99] Susceptibility and MRI simulation [Yoder-02-04] 4/32

3 2. Simulator overview 2.1. SIMRI overview 2.2. Virtual object description 2.3. Static field and field inhomogeneity 2.4. MRI sequence 2.5. Magnetization kernel 2.6. T2* effect 2.7. Noise and filtering 5/ SIMRI overview Noise Virtual object (i,ρ,t 1,T 2 ) Magnetization computation kernel k space (RF signals) B 0 +( B 0 map) MRI Sequence RF Pulse Gradient Precession Acquisition Filtering Reconstruction algorithm MR image 6/32

4 2.2. Virtual object description 3D volume of physical parameters ρ, proton density T1, T2, spin relaxation constant 1 to N components (Chem. Shift, partial volume effect) Object dimension, component resonance frequency B s : susceptibility associated field Synthetic objects perfectly known Sphere, ellipse, cube Adapted McGill brain phantom Real images segmentation + (T1,T2 maps) objects Virtual object (i,ρ,t 1,T 2 ) 7/ Static field and field inhomogeneity r r r r r r B ) z = B ( ) z + B ( ) z ( s 0 r B s (r ) B ( r ) linked to the susceptibility variation linked to the main field inhomogeneity B i 0 r linked to the intra-voxel inhomogeneity, It induces a T2* FID weighting e γ B i t T 1 * 2 1 = + γ T 2 B i 8/32

5 2.4. MRI sequence 4 types of event within C language functions Free precession (duration) Precession + gradient (x,y,z) RF pulse +- gradient - Constant pulse (duration, flip angle, rotation axis) - Sinc shaped pulse (duration, number of lobes, number of point) - User shaped pulse (file, constant pulse list) 1D signal acquisition step (number of points, bandwidth, readout gradient, k space position) 9/ Magnetization kernel Do Pulse Do Gradient M r Do Waiting (t) M r Magnetic ( t 1 ) event RF acquisition k-space filling M r ( t 2 ) Kernel is based on the Bloch Equation r M x / T2 dm r r = γ.( M B) M y / T2 dt ( M z M 0 ) / T1 and on their discrete time solution r r r r M (, t + t) = Rot ( θ ). Rot ( θ ). R. R. M (, t) z g z i relax RF 10/32

6 2.6. T2* effect, limited spin number Lorenzian distribution of isochromats 1 1 FID weighting e γ B i t = + γ T * T 2 2 B i Spin refocusing, Spin Echo α1 TE/2 α2 TE/2 exp(- B τ) i a) b) c) τ=0 τ=-τ (τ=+t) τ (τ=+t) τ (τ=+t) τ t 11/ Noise and filtering Thermal noise White Gaussian noise in the k-space K-space filtering Hamming (or any digital filter), prevent ringing Reconstruction FFT 12/32

7 3. Simulation results 3.1. Echo train 3.2. Contrast in GE and SE imaging 3.3. True Fisp imaging 3.4. Chemical shift artifact 3.5. Susceptibility artifact 13/ Echo train and T2* M xy 3000 T2 M xy T* 2 T t(s) te/2 te Gx te/2 te t(s) Simulated signal obtained after a CPMG sequence. It is composed of a train of spin echoes T2 weighted. The intra-voxel inhomogeneity is set to 10-6 T and the main field to 1 T. Simulated signal obtained after a gradient echo pulse sequence. It is composed of a train of T2* weighted gradient echoes. The intra-voxel inhomogeneity is set to 10-6 T and the main field to 1 T. 14/32

8 3.2. Contrast in SE and GE imaging SE sequence GE sequence 15/32 McGill Brain phantom 10 tissues profiles 16/32

9 3.3. True Fisp imaging 17/ Chemical shift artifact 2 π f TE = (2k + 1)π Water and fat signals in phase opposition! 18/32

10 3.5. Susceptibility artifact Geometric distortions + Intensity loss 19/32 4. Simulator implementation 3.1. Code organization 3.2. Sequence programming 3.3. Acquisition programming D interactive simulation 3.5. Distributed implementation 20/32

11 3.1. Code organization ANSI C language Running on Linux and windows OS Independent module organization Linked in a dll wrapped for being used with python for the 1D interface Parallelized using MPI to run on grid, cluster, multiprocessors. 21/ Sequence programming 22/32

12 3.3. Acquisition programming 23/ D interactive simulation 24/32

13 The Spin Player 25/ Distributed implementation Time is the enemy! Simulation time T Texc + Tacq = c X. Y. Z).[( M. N. P) + N ] s.( ex 2D Image (NxN) -Nx2 time x = 2.3 days - High resolution : small cluster 3D image(nxnxn) -Nx2 time x = 9.3 days / = 104 years! - High resolution large scale data grid 26/32

14 A Divide & Conquer parallelisation scheme Based on the free library MPI, transparent at a user level Allowing run on single PC, PC cluster, grid architecture, massively parallel machine Simulation web portal Work done in the context of European grid projects - DATAGRID Project ( ) - EGEE Project ( ) 27/32 5. Conlusion An operational MRI simulator 1D-2D-3D Sequences (SE, GE, Turbo, TFisp, STIR) Artifacts : Chemical shift, susceptibility, static field B 0, T2*, Off-On resonance Versatile, Parallelized, Interactive and GPL! 28/32

15 6. Perspectives Main perspectives Anatomic object design Sequence tuning Image processing evaluation Molecular imaging susceptibility and Cell. Con. agent New sequences / Interface with ODIN project RF / Antennas Artifact correction Flow, Diffusion & Perfusion? 29/32 30/32

16 B s Object ( ki ) Field map estimation Boundary element [Yoder02] Acquisition [Kanayama_96] - 2 EG phase images with TE 1, TE 2 - Φ + unwrap [Jenkinson_03] > B 0 r B s (r) SIMRI Boundary Integral [Balac04] B (P) NbFace 1 = Bo Xm(P) δω(p) Xm n Bo C(P, q)dsq 4π k= 1 Face k 31/32 t+ t r r θ = γ.. G( τ ) dτ g r θ = γ. B( ). t i t gradient effect Inhomogeneity effect cosθ sinθ Rot z ( θ ) = sinθ cosθ R relax t r T2 ( e = 0 0 ) e 0 t r T2 ( ) e t r T1 ( ) Relaxation effect R RF = z x z Rot ( φ). Rot ( α). Rot ( φ ) Pulse effect α' = τ. ( ω) 2 α + τ 2 Off-resonance effect 32/32

17 K-space acquisition - One point is obtained by summation all over the object r r r s[ t] = M (, t). x + j r r r r r M (, t). y - Next point is obtained after a time step t, the sampling rate 33/ Distributed implementation Time is the enemy! Simulation time T Texc + Tacq = c X. Y. Z).[( M. N. P) + N ] s.( ex 2D Image (NxN) -Nx2 time x = 2.3 days - High resolution : small cluster 3D image(nxnxn) -Nx2 time x = 9.3 days / = 104 years! - High resolution large scale data grid 34/32

18 A Divide & Conquer parallelisation scheme Based on the free library MPI Transparent at a user level Allowing run on single PC, PC cluster, grid architecture, massively parallel machine Virtual object portions RF signal contributions Virtual object 1 Seq. Computing node N 1 s 1 MRI Seq. Master node i N-1 Seq. Computing node N i Computing node N N Seq. s i s N-1 Rec. Σ FFT k space Master node MRI image 35/32 Some test results : - PC cluster CREATIS : 8 (PIII-1Ghz) + 10 (PIV-2,6 Ghz) - CINES parallel machine : 64 to 128 nodes, RI Mhz - IN2P3 through EGEE grid interface : 9 AMD Opteron 2.2 Ghz : 8h30 on a 128 proc. CINES machine Work done in the context of European grid projects - DATAGRID Project ( ) - EGEE Project ( ) 36/32

19 37/32