Throughout the history of MRI a long sequences of techniques have been developed to make images with less data. When MRI first emerged as a clinical imaging modality, images were acquired by fully sampling a rectangular region in spatial-frequency, or k-space. Faster acquisitions have been based on the structure of the data, knowledge of the receive array sensitivity, models of the statistics of the image, and, more recently, models of the images themselves. The common thread is using other knowledge we have about the MRI system and subject to provide additional prior information to reduce the amount of data we need to collect. This talk will highlight this sequence of methods, and look at the potential and challenges they present for clinical application.
Parallel Imaging (PI)
Originally clinical MRI systems used a single volume coil for the anatomy of interest, such as a head coil or body coil. Parallel imaging was enabled by the use of arrays of small receive coils, each of which sees a small part of the total imaging volume. These were simply combined to produce higher SNR images since the individual coils had smaller noise volumes. They also have a localized sensitivity which provides some localization. Parallel imaging uses this to acquire much smaller FOV images, where only a half or third of the k-space lines are collected. The conventional Fourier transform reconstruction produces aliased images for each coil. The PI reconstruction then uses knowledge of the coil sensitivities to resolve the aliased signal, and produce alias-free full field of view images.
Different parallel imaging methods are characterized by how they exploit the coil sensitivities. One fills in missing k-space data (GRAPPA). Another uses measured coil sensitivities to sort of the aliasing in image space (SENSE). There are also hybrid versions, as well as ESPIRiT which combines features of both SENSE and GRAPPA. Fundamentally, they all solve the same problem. The strengths and weakness of each are primarily due to practical implementation issues.
Parallel imaging reduces the scan time, so this reduces the SNR by the square root of the scan time. In addition, the coil sensitivities are imperfect localization functions, so this too contributes to a reduction in SNR (the g-factor). This limits PI acceleration for a 2D image to 2-3, and completely fails at a factor of 4 or more.
Simultaneous Multislice (SMS)
In 2D PI is limited by slowly varying receive coil sensitivities. However, if we have two slices that are widely separated, the receive coils for each may be completely separate. They can image at the same time with no cost in SNR. In practice there will be many slices (as many as 10 or more). If the coil sensitivies are very different from one slice to the next PI works well. We get the usual factor of 2-3 for each slice, and then another factor of 3-5 across the slices, for total factors of 6-15.
There are many variations of SMS. These vary in how they make the other slices in the set look different. Most commonly phase cycles are used to displace slices in the phase encode direction (CAIPIRINHA, or CAIPI for short). A single coil's sensitivity will appear in different locations in the different slices, making the parallel imaging reconstruction much better behaved.
Compressed Sensing (CS)
Another approach for reducing the amount of data uses the fact that MRI images are compressible. A high fidelity representation of an image can be obtained with only 5% of the data with the wavelet compression. CS aims to only collect the data needed, rather than throw it out later.
The key idea is to "randomly" undersample the k-space data, in a way such that the reconstruction artifacts look like noise rather than an image. Compressing this image using wavelets, the image component will compress well (a few large coefficients), and the reconstruction artifacts will not (many small coefficients). Zeroing out the small coefficients will mostly suppress artifacts. The reconstruction is done as an iterative optimization to find the image that is both compressible, and is consistent with the actual spatial frequency data that was acquired.
One of the challenges with the acceptance of CS is that fact that the image artifacts are different from conventional Fourier acquisitions, where artifacts such as blurring and Gibbs ringing are well known and understood. CS has its own unique characteristics that must be learned, such as the loss of low contrast structures at low SNR's.
Machine Learning (ML)
Machine learning is rapidly emerging as a powerful tool for a broad range of problems. It is having a tremendous impact on MRI research, although it is too early to say how it will be adopted in practice. It can be used for solving any of the previous reconstruction problems. A network is trained by presenting it with a large number of input-output pairs, so that it can learn a network structure that will map the inputs to the outputs. Training can take a hours or days. The trained network can be extremely fast, as much as 100 times faster for the iterative CS reconstructions.
ML can greatly expand the prior information that can be exploited, to allowing higher PI and CS accelerations. The challenge is to make sure that the images accurately reflect reality. If a network is trained exclusively on 2D axial T2 weighted head images, that is what it will produce no matter what data it is presented with. Choosing the training data so as not to overly constrain the reconstruction will be critical. ML can also be used to exploit other information, such as reconstructing a low SNR ASL image using prior information from T1 and T2 images that have been previously acquired.