Parallel Imaging and Computational MRI Mastery Guide for MRI

Foundations of MRI Speed and Array Coil Technology

Magnetic Resonance Imaging (MRI) is inherently a slow process due to the nature of data acquisition. The fundamental pulse sequence involves an RF pulse followed by Gradient Selection (Gss), Phase Encoding Gradient (GPE), and Frequency Encoding Gradient (GFE) sequences culminating in signal acquisition. To address this slowness, array coils utilize a multi-channel receiver architecture. These array coils consist of multiple coil elements (labeled Element 1 through Element 12) connected via a multiplexer and connector to various RF channels (A, B, C, D).

The implementation of array coils offers several advantages: improved Signal-to-Noise Ratio (SNR), the ability to use lower Number of Excitations (NEX), higher spatial resolution, and significantly shortened scan times. This efficiency leads to better breath-hold capabilities, greater anatomical coverage, and a reduced need to move the patient during the examination. However, each element in the array generates its own noise. Every element also possesses its own individual k-space. The composite image from these multiple elements ($I_1, I_2, I_3, I_4$) is typically reconstructed using a sum-of-squares approach, represented by the formula: $I_{comp} = \sqrt{I_1^2 + I_2^2 + I_3^2 + I_4^2}$.

Parallel Imaging in Image Space: SENSE and ASSET

Parallel Imaging (PI) in image space was significantly advanced by Pruessman et al. in 1999 with the introduction of SENSE (SENSitivity Encoding). The relationship between k-space and image space is defined by the field of view ($FOV$) and sampling intervals: $FOV = \frac{1}{\Delta k}$ and the pixel size $\Delta x = \frac{1}{FOV_k}$. When k-space is undersampled (for example, by doubling the phase encoding step to $2\Delta k$), the phase encoding field of view is halved ($FOV_{PE}/2$), leading to aliasing.

SENSE and ASSET (Array Spatial Sensitivity Encoding Technique) work in image space to reduce the number of required phase encodes. They utilize the known sensitivity maps of array coil elements to unfold the aliased data. This method is applicable to any pulse sequence and speeds up the scan according to the formula: $\text{Scan Time} = \frac{TR \times NPE \times NEX}{ETL \times \text{Reduction factor}}$.

The mathematical basis for SENSE involves calculating the signal at a point $P$ in a reduced FOV image. If $a$ is the displacement, the observed intensity $I$ is a combination of the true signal at the original position $S(y)$ and the displaced signal $S(y+a)$. For two coils ($C_1, C_2$), the equations are: $I_1 = C_1(y)S(y) + C_1(y+a)S(y+a)$ and $I_2 = C_2(y)S(y) + C_2(y+a)S(y+a)$. In these equations, the component $C(y+a)S(y+a)$ represents the displaced signal multiplied by the coil response for that displaced position. The reconstruction process requires a calibration step using sensitivity maps for each coil (Coil A, Coil B) followed by the actual acquisition.

Parallel Imaging in k-Space: SMASH, GRAPPA, and ARC

Parallel Imaging in k-space began with SMASH (SiMultaneous Acquisition of Spherical Harmonics) as presented by Sodickson & Manning in 1997. SMASH works by using a reduced number of phase encodes and employing array coil element sensitivities to create "virtual phase encoding." Unlike SENSE, basic SMASH does not require a separate sensitivity map because auto-calibration data is often embedded within the scan. The scan time reduction formula is identical to SENSE: $\text{Scan Time} = \frac{TR \times NPE \times NEX}{ETL \times \text{Reduction factor}}$.

SMASH achieves virtual phase encoding through the linear combination of coil sensitivity profiles. By combining signals from Coil 1 ($C_1$) and Coil 2 ($C_2$) (e.g., $C_1 + C_2$ or $C_1 - C_2$), different spatial harmonics are generated that mimic the effect of phase encoding gradient steps ($G_{PE}$). This allows for the calculation of missing lines in k-space. For an acceleration factor $R=3$, one line is acquired while two are calculated.

GRAPPA (GeneRalized Autocalibrating Partially Parallel Acquisitions), introduced by Griswold et al. in 2001, and ARC (Autocalibrating Reconstruction for Cartesian imaging) represent advanced k-space methods. These use Auto-Calibration Signal (ACS) lines, which are extra lines acquired integrated into the main acquisition. This makes the reconstruction less sensitive to motion. In GRAPPA, multiple lines from every coil are used to fill the k-space for each element independently. These are then reconstructed for each element before a Sum of Squares (SOS) image combination is performed. In Echo Planar Imaging (EPI), parallel imaging increases phase encoding bandwidth, which shortens the Echo Time ($TE$) and reduces distortions (e.g., $TE = 1/bwp$).

Clinical Implementation and Manufacturer Acronyms

Parallel imaging is marketed under various names by different manufacturers. Philips uses SENSE, while Siemens employs SENSE, GRAPPA, and CAIPRINHA. GE uses ASSET and ARC. Toshiba uses SPEEDER, and Hitachi provides RAPID. Specialized techniques include SPACE-RIP (Sensitivity Profiles from an Array of Coils for Encoding and Reconstruction In Parallel) by Kyriakos et al. (2000), GEM (General Encoding Matrix) by Sodickson & McKenzie (2001), and Auto-SMASH by Jakob et al. (1998).

When choosing a method, trade-offs must be considered. SENSE and GEM are generally better for tight FOV applications, while ARC and GRAPPA offer better tolerance to motion. The primary trade-off is between the acceleration factor ($R$) and image quality. The Signal-to-Noise Ratio in parallel imaging is governed by the formula: $SNR \propto \frac{1}{g\sqrt{R}}$, where $g$ is the geometric factor (g > 1). This factor represents the noise amplification due to the geometry of the coil array. As $R$ increases (e.g., from 1.5 to 4), the g-factor map shows increased noise, and foldover artifacts may appear if the reconstruction cannot successfully unwrap the aliased images.

In clinical practice, PI benefits include significantly reduced EPI distortion, reduced TE in Diffusion Weighted Imaging (DWI), shorter breath-holds in abdominal imaging, and reduced blurring in Single-Shot Fast Spin Echo (SS-FSE) sequences.

Advanced k-space Speed-up Methods and Missing Data Simulation

Beyond parallel imaging, other techniques exist to speed up data acquisition by simulating missing data in k-space. Rectangular FOV and Scan Percentage only sample a proportion of the phase encoding lines ($k_y$), which usually results in less resolution but higher SNR. Half Fourier (Partial Fourier) acquires only half of k-space and uses the conjugate symmetry of k-space to fill the rest; this maintains resolution but reduces SNR and cuts time by approximately 50%. Partial Echo reduces sampling time during frequency encoding ($FE$), resulting in a shorter TE.

Radial Sampling involves acquiring lines through different directions in k-space and regridding them into Cartesian coordinates ($k_x, k_y$). This provides better contrast and fewer movement artifacts but is not necessarily faster. Time-based data sharing, such as 3D TWIST, is used in dynamic scans to increase frame rate. The first scan is a complete k-space acquisition, while subsequent scans only repeat the central part (25-50%) and cycle through fractions of the outer parts. Missing data is estimated using formulas such as:

1. Zero filling: $S(k_{FE}, k_{PE}) = 0$
2. Complex conjugation: $S(k_{FE}, k_{PE}) = S^*(-k_{FE}, -k_{PE})$
3. Linear Interpolation: $S(k_{FE}, k_{PE}) = \frac{S(k_{FE}, k_{PE+1}) + S(k_{FE}, k_{PE-1})}{2}$

Computational MRI: SMS and Multi-Band Imaging

Simultaneous Multi-Slice (SMS), also known as Multi-Band Imaging, is a modern computational approach that acquires multiple slices simultaneously. Data is dissected into individual slices using combinations of Phase Encoding, Frequency Encoding, RF Phase Offsetting, and coil sensitivity.

The advantages of SMS include reduced scan times with zero penalty for SNR, which theoretically makes it superior to standard Parallel Imaging. In EPI, it reduces TE and subsequent distortion. However, the disadvantages include higher RF power requirements (increased SAR), dependence on high coil channel counts and geometry, and the necessity for large slice gaps, as it is difficult to separate adjacent slices.

Compressed Sensing (CS) in MRI

Compressed Sensing (CS) is an image processing approach adapted for MRI, similar to JPEG compression. It relies on the redundancy of data to allow for undersampling. CS requires three components: a sparse image representation (often achieved via a Wavelet transform), an incoherent (quasi-random) sampling strategy, and a non-linear (iterative) reconstruction process.

Because MRI data is collected in lines, radial or spiral sampling is often used to achieve the necessary incoherence. Iteration is used to reconstruct the data using "L-1 norm minimization," which fills in missing data by retaining sparsity across different scales. This provides a "best guess" that is iteratively improved. The final images typically differ from full-sampled images only in the structure of the noise. The primary advantages are reduced scan times with minimal SNR loss and lower RF/SAR. The disadvantages include changed contrast, different artifact appearances, and the need for intensive post-processing and specialized training for radiologists to interpret the results correctly.

Magnetic Resonance Fingerprinting (MRF)

MR Fingerprinting (MRF) shifts the focus from direct imaging to analyzing tissue properties. It uses randomly varying acquisition parameters (TE, TR, Flip Angle) to acquire incoherent data. The signal evolution of each pixel is then matched against a "dictionary" or lookup table of known tissue types (T1, T2, PD, etc.).

Once a match is found, "virtual" images can be reconstructed with any desired contrast (T1, T2, PD, FLAIR, DWI). This technique typically produces quantitative T1 or T2 maps rather than just weighted images and can be combined with Parallel Imaging. MRF has the potential to massively speed up examinations by replacing multiple sequences with one acquisition. However, the data is essentially "artificial," reliant on the accuracy of the dictionary, and currently has limited clinical application. It represents a transition from qualitative weighted imaging to quantitative tissue mapping.