What is Multi-Planar Reconstruction ? (MPR)
What Can MPR be used for ?
MPR Image Quality Issues
3DFT Acquisition Principles
3D Raw Data & K-Space
Non-Selective Vs. Selective 3D
RF Pulses & Slice Profile
Manipulating 3DFT SNR
Artefacts in 3DFT
3DFT Data Handling
3DFT Sequence Types
Isotropic T1 weighted Brain Imaging
Siemens Solution T1 MP-RAGE
GE Solution IRprepped FSPGR
High Resolution Brainstem Imaging
|Back To Adelaide MRI Website||Contact Author by e-mail|
While there is nothing conceptually new about 3DFT and MPR, recent developments of commercial hardware capability and sequences now make them practical and easy to implement. In MRI now is always a good time for change. This paper aims to get you to make that change.
The installation of a replacement MRI scanner at the Royal Adelaide
Hospital offered the opportunity to review and adapt virtually all aspects
of MRI technique. Fast Spin Echo (FSE or TSE) phased array coils, and stronger
gradient systems allowed substantial improvement in image resolution, particularly
in spine, joint and T2 weighted brain imaging. In seeking a solution to
thinner slices and higher resolution for T1 weighted brain imaging, 3DFT
sequences and Multi-planar Reconstruction (MPR) were employed. Once the
step of using 3D sequences and MPR was taken, we found a range of practical
3DFT sequences suitable for exploiting MPR as a daily standard technique
tool in clinical applications. It is particularly applied where more conventional
approaches have reached their limits of speed or resolution, where the
anatomy under examination requires complex and curved plane display, or
where multiple plane examinations make 2D acquisition inefficient.
What is Multi-Planar Reconstruction?
Multi-planar reconstruction (MPR) software re-formats images at arbitrary
planes, defined by the operator, using the pixel data from a stack of planar
images (base images).
The digital value for each pixel is assigned to a virtual voxel with the same thickness as the slice thickness of the base image. This yields a volume of data that represents the scanned object. The MPR software uses the virtual voxel data to create the pixel values for the reconstructed images. When the dimensions of the scanned voxels (as set by slice thickness and in-plane resolution) are equal, the data set is said to be isotropic. Where the dimensions of the scanned voxel are not equal in all three planes, the data set is said to be anisotropic. Isotropic data yields the best reconstructions.
When applied to CT and MRI data the MPR images can aid perception of anatomy by providing a perspective or display not seen in the base images.
MPR software typically uses an interactive interface that allows the user to prescribe the reconstruction planes and parameters from simple reconstructed images, in a manner analogous to scanning the real patient. The user can select plane orientation and thickness (typically equal or greater than the base scan thickness), and can prescribe the number, location, and separation of reconstructed slices. Individual MPR programmes have their own sets of rules and constraints for data input which can at times place limitations on the parameter choices in the base sequences. Of greatest importance is the number of data points allowable (slices* matrix), which will largely be a function of computer memory. This limit is rarely expressed explicitly but will be encountered with adventurous sequence development. Most MPR software can accept 2D slice data, and can accommodate some slice gaps.
What can MPR be used for?
Once suitable base scans have been performed, we have a tool to "virtually"
scan the patient without further need to keep them in the scanner. The
base scan data can be used to create images precisely aligned with anatomical
planes, create sets of slices with multiple orientations and slice thicknesses
(at the same contrast weighting as the base scans), interactively roam
through anatomy, and compensate for positioning errors.
The MPR technique is an accessible tool that can replace or complement more conventional imaging approaches. MPR can be routinely used in examinations of the CP angles and cranial nerves, pituitary gland, solitary and multiple space occupying lesions of the brain, in the knee joint, and for examining the kypho-scoliotic spine.
The quality and utility of MPR images is directly related to the quality
of the base images used to create them. In this sense the factors affecting
image quality are the same as for any scan. Suitable contrast, SNR, contrast
to noise ratio (CNR) and the minimisation of artefacts are most important.
Because most MPR work involves the use of short TR GRE 3DFT sequences,
obtaining suitable image contrast can be difficult. As will be discussed
later, the best solutions are offered by complex and efficient GRE sequences.
Geometric parameters assume an added importance in this application. Slice thickness cannot be used to provide SNR to compensate for losses inherent in providing high in-plane resolution. Optimum results require slice thickness equal to the in-plane pixel size (isotropic data). When the data is not isotropic, images reformatted using the larger dimension will show reduced spatial resolution. This problem is worst when the large dimension is in the reconstruction plane, but its visibility will depend on the level of anatomical detail that needs to be resolved.
Artefacts causing geometric distortion of the base image will also degrade the MPR image fidelity. Contrast and signal loss from cross talk becomes a problem when designing thin slice contiguous 2DFT sequences, particularly in multi-slice sequences with low acquisition bandwidths. While attempting to maintain a reasonable number of slices per unit TR, the sequence designer may choose a shorter period RF pulse, to make up for the increased sampling time dictated by a low acquisition bandwidth, sacrificing pulse shaping and slice profile in the process.
The MPR software will employ specific algorithms to assign the pixel values of a reformatted slice. The image quality will depend on how well these handle multiple voxel values in thick slices, and how the data gaps resulting from inter-slice gaps are interpolated. The interpolation routines of magnification programmes can also significantly affect presented image quality.
In any given system best results are obtained with small dimension isotropic data sets acquired from 3DFT data. When using 2DFT sequences, keep slice thickness and slice gap as small as possible, and plan to reconstruct planes at small angles to the acquired plane.
For best results use (in order of preference):
3DFT Image Acquisition
3DFT is a spatial encoding regime that can be applied to virtually any excitation sequence. It is also referred to as 3D, or Volume acquisition. 3DFT sequences still generate two dimensional images in the acquisition plane; they differ from 2D acquired sequences in that the MR signal is generated from the entire imaged volume, rather than the individual imaged slices. For any given image thickness, the 3D sequences use lower gradient strengths than 2DFT.
Figure 2 shows a pulse-timing diagram for a slice selective GRE 3DFT sequence.
The volume of tissue to be imaged can be specified by using the same technique as slice selective 2D imaging, ie. an RF pulse with a controlled bandwidth is applied in conjunction with a slice select gradient (Gz-ss). This tissue volume is resolved into slices by applying a phase encoding gradient in the slice select direction (Gz-phase). The number of increments of Gz-phase equals the number of required partitions. The excited volume is called the SLAB or Slice, while the resolved slices are called PARTITIONS or Sub-Slices.
Usually Gz-phase is incremented through its full range of values while the Gy phase encoding gradient is held constant. The process is then repeated with the next increment of Gy phase.
For conventional sequences, 3DFT sequence time = .
Incrementing Phase Encoding Gradients
The phase encoding gradient must be applied with a range of discrete values from
-Gmax to +Gmax to collect the required range of raw data. The number of steps determines the resolution in the phase encoding direction concerned. The maximum gradient strength (Gmax) determines the field of view in that direction. Low strength gradients, called the central steps (close to G=0), control the contrast of an image while the higher strength steps (+Gmax & -Gmax) are called the outside steps and determine the final resolution of the image. These titles refer to the locations in K-space that the phase encoding steps provide.
Sequence designers can employ a range of patterns to apply each of the
gradient steps. These patterns are called the phase ordering. In sequential
phase ordering, the gradient begins at -Gmax, stepping through G=0
to +Gmax (or vice versa). In centric phase ordering the central
phase encoding steps are acquired first, progressing to the higher gradient
steps in a balanced manner (+1,-1, +2,-2,+3,
-3....+Gmax, -Gmax). Other incrementation strategies are referred to as free phase ordering. A wide range of free phase ordering regimes have been designed to meet specific needs, of which the most common deal with respiratory or cardiac motion compensation.
Raw Data & K-Space
3DFT sequences generate three-dimensional raw data (Kx, Kz, Ky). It can be reasonably assumed that the nature of 3D K-space is similar to 2D K-space data. K-space trajectories are controlled by the spatial encoding regimes, particularly the phase ordering strategies, and the regions of K-space near the origin (low values of Kz, Ky) contribute most to image contrast. The data lines at higher values of Kz contribute spatial resolution of partitions, and the higher values of Ky contribute to in-plane resolution.
Non-Selective & Selective 3DFT
Some 3DFT sequences will operate with "hard" or non-selective RF pulses, in order to minimise TR. Others will be designed with selective RF pulses at the expense of minimum TR or bandwidth. When using a non-selective pulse the entire object should be included in the slab thickness to avoid aliasing of signals from outside the slab. A small FOV coil may also prevent this artefact.
When the TR must be long to provide the required contrast weighting multiple slabs can be excited in the same way that multiple slices are usually excited in 2DFT sequences.
Manipulating 3DFT SNR
Because the signal is derived from the slab not the slice, the SNR of a 3DFT is significantly higher that a comparable 2DFT for a given volume of tissue. If we compare a multi-slice 2DFT sequence with a 3DFT sequence offering the same number of slices and thickness:-
In non-selective excitation, the slice select gradient isn't applied
so the entire patient resonates at a single frequency, and a fast broadband
RF pulse will excite all tissues within in the transmit coil. For slice
selective excitation the SS gradient dictates that the resonant frequency
will vary continuously along the slice direction. To achieve a rectangular
slice profile the RF pulse must contain only those frequencies that equal
the Larmor frequencies of the tissue in the selected slice location and
thickness. Uniform excitation of the slice requires equal amplitude for
all frequencies in the excitation bandwidth. In theory a pulse which contains
a specific range of frequencies at equal amplitude would be infinitely
long. In practice reasonable slice profiles are achieved with a pulse period
of 8-16 cycles, but the longer the better.
Therefore the broadband excitation pulse suitable for non-selective excitation can be applied in less time than a well controlled narrow band pulse needed to achieve selective excitation.
Artefacts in 3DFT
Any artefact that can affect a 2D image can affect a 3D image in the same way, although they can appear differently due to the difference in spatial localisation method. The phase encoding for partitions is prone to the same motion ghosts as normal in-plane phase encoding, but they will extend in the slice select direction and be noted on adjacent images. These can be seen across the slice select direction so the ghosts of the pulsatile object may not appear in the same slice as the artefact, nor in every slice. In multi-slab 3DFT, ghosts in the slice select direction will be restricted to the slab containing the source.
Aliasing is common in the slice select direction whenever the object is larger than the slab width. The signals from outside one edge of the slab will wrap around to the slices at the opposite edge of the slab. This is commonly seen as extra ears on sagittal head slabs, or bilateral display of the fibulae in knee sequences. Restricted view coils can minimise this aliasing. Phase over-sampling in the slice select direction is more effective but costs time.
The profile of the slab is prone to the same distortions as 2DFT slice profiles, and the poor slice profile common to many short TR selective 3DFT sequences is frequently seen. Edge slices will appear with poor signal level and little contrast. The severity of the effect depends entirely on sequence design and needs to be assessed individually. Its appearance is often compounded by slice direction aliasing. It is overcome by extending the thickness of the slab by 10-30%, acquiring extra partitions and discarding the poor images.
The slice profile of a partition is rectangular, so there is no loss of contrast due to cross excitation (cross-talk) between partitions.
3DFT Data Handling
Reconstruction times are longer than 2DFT sequences due to the extra Fourier Transform step. Actual reconstruction times will depend on the number of voxels in the 3D raw data set, and the Fourier transform algorithms supplied by the manufacturer. In older scanners the Fourier transformation algorithms work more efficiently when the number of phase encoding steps and partitions is a power of 2 (2,4,16,32,64,128,256,512). Selection options may reflect this.
3DFT sequences generate very large raw data sets. When using phased array coils the raw data volume is multiplied by the number of coil elements, and can exceed the data handling capacity of the scanner. This restricts some applications at present.
Handling large amounts of raw and image data efficiently requires high computer memory, multi-image screen display, and fast archive and image recall software. Current release scanners can handle the data from most 3DFT applications adequately although there is need for improvement.
At the Royal Adelaide Hospital, we currently archive the base data as well as all the reconstructed slices so that the images used to report the examination are available as well as the capacity to create further views.
3DFT Sequence Types
The major challenge for 3DFT sequences is scan time. For whole object isotropic data sets approximately 160 partitions are required with about 200 phase encoding steps so scan time = TR*(32,000 to 43,000 mSec). To obtain scan times less than 10 minutes, TR must be kept short. As TR falls below 100 mSec low flip SE and GRE sequences generate very low signal levels and have a poor range of image contrasts. More efficient sequences such as Contrast prepared GRE, segmented K-space GRE, GRE, PSIF, CISS, and DESS offer practical combinations of contrast, SNR and scan time.
Isotropic T1 Brain Imaging
Problem: Thin T1 images
|Thickness||5 mm||4 mm||3 mm||2 mm||1 mm|
|SNR drop per change||20%||25%||33%||50%|
|NEX for equal SNR||2||3||6||13||50|
|Scan Time (minutes)||3:12||4:48||9:36||20:48||80:00|
T1 weighted TSE scans using low turbo factors (ETL=3-5) display poor Grey/White matter contrast. T2 decay induced blurring and ghosting are also significant problems. Turbo IR (FIR) sequences can use higher ETL (7-9) to yield good T1 weighting, but they offer limited slices and use very long TR. Their application is limited by scan time when very thin slices are required.
Conventional spoiled GRE can reduce TR by a factor of 5, to about 100 mSec, but the possible number of slices per unit TR demand multiple sequences (packets or packages) to achieve coverage. 3DFT spoiled GRE with TR near 100 mSec, overcomes the slice limits, but the scan times are still very long.
Low bandwidth sequences used to compensate low SNR, require longer data
acquisition times, reducing the number of slices per unit TR. Attempts
to recover this lost time by reducing the period of the RF pulse result
in poor slice profiles and more cross-talk. (see RF
Pulse Shapes & Slice Profiles)
Siemens Solution: T1 weight MP-RAGE
MP-RAGE stands for Magnetisation Prepared RApid Gradient Echo. MP-RAGE is a 3DFT contrast prepared gradient echo sequence. The T1 weighted MP-RAGE uses a 1800 inversion pulse followed by a time delay (TI) to establish T1 contrast in the same manner as an Inversion Recovery (IR) sequence. As Mz evolves, the signal is acquired by a spoiled GRE sequence (Turbo-Flash) with a low flip angle and extremely short TR.
MP-RAGE uses sequential ordering of the in-plane (Gy) and slice select (Gz-phase) phase encoding. All Gz-phase encoding lines are collected following the contrast preparation pulse. The process is then repeated for the next value of the Gy phase encoding gradient. This strategy gives the shortest scan time as the number of partitions (Gz-phase steps) is usually smaller than the number of Gy phase encoding steps. Each Kyz line of data has a different contrast weighting as it is collected at a different period after the inversion pulse. The sequence contrast is dictated by the effective Inversion Time (TIeff) which is the time elapsed between the inversion pulse and the collection of the central Gz-phase steps (Ky fixed, Kx changing, Kz near 0). The 2DFT implementation of MR-RAGE is called contrast prepared Turbo-Flash.
Isotropic T1 MP-RAGE
The Siemens suggested protocol for MP-RAGE has been modified at many
sites to provide isotropic 1 mm resolution. In the versions we currently
use, the inversion and excitation pulses are not slice selective so the
whole object must be included in the field of view. MP-RAGE sequences are
available with longer TR and slice selective pulses, but in practice the
sequence times are virtually equal for a given pixel volume. (See
Manipulating 3DFT SNR). The sequence exhibits strong T1 weighting
which is best appreciated on narrow window setting (use slow window control).
Its high bandwidth masks susceptibility effects and keeps artefacts from
dental and other metal very small. There is significant flow related enhancement
in the inferior 2/3 of the images which allows MIP MRA images of the carotids
to about the trunk of the middle cerebral artery. This drop-off of signal
is due to progressive saturation of blood as it stays in the slab for successive
excitations, and so the display of vessels would be better post contrast.
Bright and dark pulsatile ghost artefacts result from flow in the internal
carotid and basilar arteries. They are seen typically in the brainstem
and cerebellum although not only in the sagittal partitions that contain
the vessels. Eyeball motion is best controlled when the patient is scanned
with their eyes closed. The sequence has good SNR so it can handle a 6/8
rectangular matrix. AP aliasing of the nose must be kept out of the posterior
fossa. If phase over-sampling is used to control aliasing with the smaller
matrix, remember that it may return the scan time to its original value.
Keep an eye on the relative SNR indicator.
Characteristics of Isotropic MP-RAGE
Water Excitation MP-RAGE
The inversion pulses and excitation pulses are frequency selective for water protons therefore fat signal is virtually eliminated with a method suitable for post contrast applications. The slight increases in TR increase the scan time marginally. Because the fat signal is reduced by spectral methods rather than inversion, the sequence is suitable for use with Gd contrast agents. It is well suited to investigation of skull base tumours and optic nerve tumours, post contrast. The data set can be used to MIP angiograms of the major neck and cranial vessels with better results than the standard MP-RAGE.
Sequence parameters for Water Excitation MP-RAGE
CISS (Constructive Interference Steady State)
CISS is a strongly T2 weighted GRE sequence. In essence it is a pair of True FISP sequences acquired with differing regimes of alternating the phase of the excitation pulses. Individually these True FISP sequences display very strong T2 weighting but are affected by dark phase dispersion bands which are caused by patient induced local field inhomogeneities and made prominent by the relatively long TR used. The different excitation pulse regimes offset these bands in the two sequences. Combining the images results in a picture free of banding. The image combination is performed automatically after data collection, adding some time to the reconstruction process.
The overwhelming power of the 3D CISS sequence is its combination of high signal levels and extremely high spatial resolution. CISS images yield the best detail available of the cisternal portions of cranial nerves. In combination with the isotropic MP-RAGE we believe they completely remove any need for contrast media in identifying acoustic neuromas.
The sequence has inherent flow compensation because of its perfectly balanced gradients. Compared to conventional FISP or GRASS it is quite insensitive to CSF pulsations. True FISP and CISS sequences require a very high level of control over gradient switching and shaping. CISS requires very high local field homogeneity so an excellent base magnet homogeneity is required, and all sequences must be preceded with a patient specific shim adjustment. Metal in the field will degrade the images substantially so patient preparation should include the removal of all head and neck jewellery, and metal from clothing. CISS is available in 2DFT and 3DFT implementations.
TR 26.7 mSec TE 9 mSec Flip 60 degrees
128 mm sagittal slab 128 partitions
Matrix 256 x 256 FOV 250 mm 5/8 Phase A-P
1 Acquisition Scan Time 9:08 Resolution 1 x 0.98 x 0.98 mm
TR 2800 mSec TE 120 mSec ETL 27 Flip 1600 - 1800
Sagittal 5 slices 8 mm Gap 100% 8 Partitions Slice select phase over-sampling 50%
Matrix 256x189 (98%) FOV 280x210 (6/8) Phase P-A Frequency over-sampling
1 Acquisition Repeat sequence with 8 mm shift
1 spatial saturation band anterior coronal
Resolution 1 x 1.11 x 1.09 mm Scan Time 3:58 (x2)
If you would like this paper presented at your MRI meeting or you have comments or suggestions, please contact the author by e-mail.