Compressed sensing Multiband MRI A new approach to

Compressed sensing & Multiband MRI A new approach to speed up MRI Vahid Shahmaei MRI MSc MRI Application Specialist Forough Sodaei Medical Imaging MSc

MRI Acceleration options q Parallel imaging (multiple receiver coils) q Partial Fourier (symmetry of k-space) q Keyhole, view sharing (energy distribution) q Compressed sensing (sparsity) or Hyper sense q Hyper band or Multiband v In many cases these can be combined

COMPRESSED SENSING Using Compressed SENSE appeared a simple yet powerful way to accelerate MRI scanning for different contrast types and sequences, in 2 D as well as 3 D Compressed SENSE works in virtually all anatomical areas and with many different scan techniques and contrast types That shorter scanning time will then benefit our patients and in addition, it will allow us to scan more patient replacing 2 D sequences with 3 D sequences What is the easiest way to shorten MRI scan time? !

COMPRESSING SENSING v Incoherent subsampling (high acquisition speed) q Fewer raw data sampled q Increase scan speed q With incoherent sub sampling avoid distinct aliasing artifact Ø reducing the data sampled in a randomized fashion results in an incoherent aliasing pattern that tends to resemble noise. Ø When a special reconstruction is used to enforce sparsity, the incoherent aliasing can often be reduced to unnoticeable levels. Incoher ent subsam pling COMPRESSE D SENSING Transform sparsity None-linear iterative reconstruction

COMPRESSING SENSING v None-linear iterative reconstruction (balanced data consistency) • The constrained reconstruction used in CS uses only a fraction of the MRI data to reconstruct an image. • Because there are many possible images that may fit the acquired data, the reconstruction constrains the final image to be compressible v Transform sparsity (remove noise) § CS instead relies on the fact that the information in the image is actually sparse, or compressible. • the information of an image can be compressed by a factor of 10 using a wavelet transform, similar to what is used in JPEG image compression. v sparse, which tends to give the correct image.

C. S Applications Dynamic imaging v Compress sensing in neuroimaging Ø Acquire 1 mm isotropic 3 D brain images in all relevant contrasts in 3 minutes each Breath hold imaging 3 D sequences Ø Decrease scan time by 40 -50% Cardiac imaging Ø Better image quality in same time Abdomen imaging Ø Motion free imaging capable Ø Perform high-resolution TOF imaging in 2: 00 minutes for time critical patients with up to 10 -fold acceleration Ø High resolution new sequences; 3 D black blood of temporal and orbit Motion free imaging FMRI

3 D FLAIR, scan time 5: 02 min, voxel size 1. 1 x 1. 1 mm, Ingenia 1. 5 T. 3 D FLAIR, scan time 3: 02 min, voxel size 1. 1 x 1. 1 mm, Ingenia 1. 5 T.

a routine 3 D TOF with Hyper Sense in a scan time of 3: 10 min the ultra-fast 3 D TOF with Hyper Sense) in a scan time of 38 sec.

COMPRESSED SENSING v Compressed sensing for breast MRI Ø Compressed sensing offers spatio-temporal resolution improvements CS allows acceleration of data acquisitions, but requires high signal-to-noise ratio to work without producing artifacts due to the reduced amount of data. acceleration of four or more, and this can be combined with CS since the two acceleration schemes use complementary mechanisms. Higher resolution with same time Breast perfusion Ø Ø v Very early phase ( 0 -60 sec after contrast injection ) § Breas vessel evaluation- A-V interval , increased vascularity § Kinetic characterization using UF-DCE MRI § Detection of enhancing lesions

COMPRESSED SENSING v Challenges of Compressing sensing Ø Reconstruction computation: Ø Artifact : § Data sixe: Matrix + Frames + coils = Large! q Blurred images § Noise! Only use when SNR is sufficient q Washed contrast § Multiple sparsity constraints add complexity § Iterative reconstruction 10 -50 x computation Ø Integration with other acceleration options § Retrospectively discard data q Lose of resolution q Loss of low contrast q Use when limited with encoding time not SNR