Real-Time MRI using NLINV

For recent applications see Real-Time MRI in the applications section. 


MRI - The Speed Problem

Spatial Encoding Image Reconstruction
Magnetic Field Gradient G Inverse Fourier Transformation

Rapid MRI - FLASH

No Waiting Periods

Low Flip Angle Gradient-Echo MRI

 

Applications

Continuous Acquisitions @0.5-2s

ECG-synchronized Cardiac MRI

MR Angiography & Susceptibility

Three-Dimensional MRI

J Frahm et al, German Patent P 3504734.8, February 12, 1985


Real-Time MRI - FLASH + NLINV

Extreme Data Undersampling

Radial Gradient-Echo MRI

Advanced Parallel Imaging (NLINV)

Temporal Regularization

 

Applications

Continuous Acquisitions @≥ 10 ms

Dynamic MRI in Real Time

Frahm, Uecker & Zhang, US Patent 8, 384, 383 B2, March 23, 2010 


Real-Time MRI - Radial Gradient-Echo MRI

Rapid Gradient-Echo Sequences

  • RF Spoiled FLASH (T1)
  • Refocused FLASH (T2/T1)
  • Fully Balanced SSFP (T2/T1)

Radial Encoding

Robust against Motion

  • No Phase Encoding
  • Central Oversampling

Equivalence of Spokes

  • Tolerant against Undersampling

Different Spokes for Serial Acquisitions

 S Zhang et al, J Magn Reson Imaging 31:101-109, 2010 [Online Version] 


Parallel MRI - Image Reconstruction by Nonlinear Inversion (NLINV)

Joint Estimation of Image & Coil Sensitivities

Spatial Encoding   Image Reconstruction 
Magnetic Field Gradient G   Inverse Fourier Transformation

M Uecker et al, Magn Reson Med 60:674-682, 2008  [Online Version] 


Real-Time MRI - NLINV with Temporal Regularization

Joint Estimation of Image & Coil Sensitivities

Minimization of Cost Function  (  =Image/Coils,  =Measured Data)

$$\Phi (x) = \frac{1}{2} ||Ax-y||_2^2 + \sum_i \lambda_i \cdot R_i(x) \qquad x = argmin_x \Phi (x)$$

Data Consistency

Temporal Regularization

M Uecker et al, NMR Biomed 23:986-994, 2010  [Online Version] 


Real-Time MRI - Accelerated Computing

Online Reconstruction with Integrated GPU-Computer

  • Parallelization of the NLINV Reconstruction Algorithm  

  • Implementation on Server with 2 × 4 Graphical Processing Units (Supermicro, NVIDIA Titan X) 

  • Development of a Unique Library for Multiple GPU Programming 

  • Full Integration into Commercial MRI System (Siemens VD13)

 

  • Up to 50 Frames per Second

Schaetz & Uecker, Lect Notes Comput Sci 7439:114-128, 2012  [Online Version] 

Real-Time MRI - Temporal Fidelity

Experimental details: RF-Spoiled Radial FLASH , 1.6 × 1.6 × 6 mm3, Acquisition Time 33.3 / 10.0 ms, 30 / 100 Frames per Second 

J Frahm et al, The Open Med Imaging J 8:1-7, 2014  [Online Version]

Go to Editor View