The essence of single-shot echo-planar imaging (EPI) is to acquire all k-space data for a 2D image following a single RF excitation pulse on the time scale of a few dozen milliseconds instead of minutes for conventional MRI. Peter Mansfield invented EPI in 1977, in an era with not-that-homogeneous 0.1T magnets, weak and slow gradient systems, and 512k RAM image-processing computers, making any form of NMR imaging challenging. Most of what we know today regarding MR pulse sequences was uncharted territory at that time, and it took years until the hardware was good enough to form a decent 32x32 image.
Despite the hardware limitations in the late 70s, the insights into EPI's future clinical potential and the advantages of EPI were, for the most part, similar to how we would describe this snapshot MRI technique today. In a paper entitled "Biological and Medical Imaging by NMR" (1), the authors describe that EPI can overcome the "slow NMR imaging process on a living person" and that apart from patient discomfort (with a long scan time), there are scientific reasons for imaging at high speed, referring to the physiological motion of organs such as the heart, stomach, and breathing. In the same paper, the idea of 3D EPI is also discussed, decades ahead of actual clinical introduction as a multi-shot SWI/T2*w 3D EPI technique.
In 1984, Mansfield (2) also made the point that EPI can reduce the need to anesthetize patients with anxiety, claustrophobia, and children and showed pediatric cardiac and body EPI data on infants. In 1987, Mansfield's group published "Real-time Movie Imaging from a Single Cardiac Cycle by NMR" (3), showing one of the first dynamic NMR scans of the heart with 20 FPS on 0.1T, stressing the importance of single-shot real-time EPI for arrhythmic heart imaging. In 2003, Mansfield got the Nobel Prize for the development of MRI and his early work on echo-planar MRI.
When it comes to emphasizing the advantages of EPI, its inventor was able to foresee many of them nearly half a century ago, which may be summarized as:
- Freezing of physiological motion, i.e. movements of organs such as the heart and lungs, thereby avoiding image blur and ghosting artifacts. Also, imaging the dynamic movements of these organs
- Freezing of patient's involuntary head motion (children, claustrophobic patients, etc.) and the less overall need for sedation
- "4D" imaging (need for many slices in space and time): many images per second, enabling e.g. today's "4D" data sets in diffusion MRI, fMRI, and perfusion MRI
Another good summary of the strengths of the EPI sequence is its high percentage of data collection ("readouts") relative to the total sequence duration. In the simplest case, a 5 ms RF excitation pulse is followed by a 50 ms long EPI train, during which the MR signal is collected more or less continuously. This amounts to spending about 90% of the scan time listening to data. Comparing this with a conventional gradient echo sequence where the readout gradient is on for maybe 30-50% of the time, or less if the echo time (TE) needs to be larger than the minimum TE. Given a total need for a certain amount of k-space data, EPI can collect it in a competitive amount of time. EPI data may appear to have low signal-to-noise (SNR) at first glance compared to conventional MRI, but a fairer comparison is to measure
$$\eta = \frac{SNR}{\sqrt{scantime}}$$
Worth noting, the scan time efficiency is also constant regardless of how many averages (NSA) are used since SNR is proportional to sqrt(NSA) and NSA linearly increases the scan time
$$\eta = \frac{SNR_{nsa}}{\sqrt{scantime_{nsa}}} = \frac{SNR \cdot \sqrt{NSA}}{\sqrt{scantime \cdot NSA}}= \frac{SNR}{\sqrt{scantime}}$$
Today, 40-50 years after Mansfield's early work, the MR hardware has improved beyond fair comparison. As the static magnetic field strength, B
0, has risen from 0.1T to 1.5-3T, the SNR has increased proportionately, enabling higher resolution in a shorter time.
In the contemporary clinical routine, the EPI sequence comes in two main flavors - without (gradient-echo EPI, GE-EPI) or with (spin-echo EPI, SE-EPI) an RF refocusing pulse between the excitation and the EPI readout train. GE-EPI is used to get fast stacks of T2*w 2D images over time ("4D") for brain applications such as functional MRI (fMRI), perfusion (rCBV, rCBF, MTT), but also just as static T2*w EPIs to detect bleeding in the brain. GE-EPI is great for T2*w (dynamic) imaging because of its speed and a good match between EPI train length and optimal TE at 3T. At 3T, the optimal TE for all three applications is around 25-30 ms (at 1.5T, the optimal TE for T2*w contrast is ~50 ms). For a common EPI readout train duration of 40-50 ms, one can place it directly after the RF excitation and reach that TE halfway in the EPI train when the center of k-space is traversed, leaving essentially no dead time in the sequence. Increasing the TE increases the blood sensitivity (which we like), reduces the signal (more T2* decay), introduces dead time between RF excitation and the EPI train, and results in fewer slices per second.
Doing T2*w/SWI 3D scans using GE-EPI is also possible, but not as a single-shot sequence. By adding a phase encoding gradient in the slice (kz) direction and playing out one kz step each TR, a 3D T2*w multi-shot GE-EPI scan with full brain coverage can be done in as short as 20 seconds, depending on the desired resolution. Also, for this acquisition, there is essentially no dead time in the sequence, but being a multi-shot 3D scan makes it as sensitive to head motion as conventional Cartesian MRI, which leads to image ghosting.
With an added RF refocusing pulse, we get a SE-EPI sequence, which is not particularly sensitive to bleeding or other T2* effects as its signal largely depends on T2 relaxation. The SE-EPI sequence produces a fast but medium-quality T2w image, which has not been very interesting for clinical or research use since it lacks the T2*w blood sensitization but is not good enough as a clinical T2w anatomical image. But when adding diffusion gradients around the RF refocusing pulse, we get a diffusion-weighted EPI sequence (DW-EPI), which has had tremendous success since the 90s. For clinical use, DW-EPI is used to detect acute infarcts and to guide the radiologist about the tissue viscosity when e.g. differing between a tumor and an abscess. For research use, mapping white matter microstructure using diffusion tractography is hard to miss in the community. Also, for DW-EPI, it is desirable to be time optimal with minimal dead time in the sequence. Introducing the RF refocusing pulse with straddling diffusion gradients combined with some finite duration of the EPI train sets timing constraints can lead to gaps in the sequence. A common way to mitigate that is to let the EPI train not cover the entire k-space in the phase encoding direction, but some 60-80% or so, and use partial-Fourier reconstruction to synthesize the missing data. This makes the effective TE in the EPI train occur earlier in the sequence, which in most cases will make the sequence have little dead times, allowing more slices per second and shorter overall scan times.
Today's B0 fields are very homogeneous compared to the late 70s, but the fact that the patient is inserted in the field introduces local anatomy-driven field inhomogeneities. As the B0 field is now 15-30 times higher than in the early days of EPI, patient-induced spatial field inhomogeneities (negligible back then) become large. These come mainly from air-tissue boundaries around e.g. the sinuses, the ear canals, intestines, etc., which make the field lines bend. For GE-EPI, this leads to large signal dropouts. For SE-EPI and DW-EPI, the signal will not disappear but be geometrically distorted near these tissue-air interfaces of the brain (and correspondingly in other parts of the body).
The next issue to be aware of with EPI is the slow traversal in k-space (compared to just reading one line), which does not only lead to geometric distortions. The chemical shift of the fat (present in the scalp and around the eyes for brain MRI) ends up moving the fat signal by ~10 cm in the phase encoding direction (unlike conventional MRI, where the chemical shift occurs by ~1-2 mm in the frequency encoding direction). This large shift in fat signal requires all uses of EPI in one way or another to remove the fat signal from the scene. This is often made with a spectral-spatial (SPSP) water-only RF excitation pulse but can also be done with a separate fat-saturation RF pulse before the normal RF excitation. Either way, these pulses add 5-10 ms to the sequence, but no other good alternative exists.
Finally, there is no ISMRM without yet another way of trying to remove Nyquist ghosting in EPI. This ghosting artifact is a faint copy of the image, shifted by half-a-FOV in the phase encoding direction, and stems from very small hardware gradient delay differences between the readout and phase encoding gradients. This delay is of no concern for conventional MRI, but as EPI plays out readout lobes with alternating polarity, the even k-space lines must be first flipped before continuing the image reconstruction. This also means that these lines are effectively time-reversed, making any hardware delays shift the odd and even lines relative to each other in the readout direction. Luckily, today's image reconstruction workflows have well-working means to correct this with or without additional calibration data, both in a clinical and research setting.