Introductory Chapter: New Advances in MRI Clinical Analysis

1. Introduction

1.1 Detecting the nuclear magnetic moment

While magnetic resonance imaging has been used clinically since the early 1980s [1], the development of the technique traces its origin to Bloch and Purcell’s demonstration of the nuclear magnetic phenomenon several decades earlier [2, 3]. Their work showed that in the presence of a constant magnetic field, nuclear magnetic moments generated a resonance frequency that could be electronically detected by means of an inducting coil. The ability to detect an electrical signal induced by the magnetic moment made possible the development of adjunct procedures for spatially profiling analytical specimens. Advances in MRI that have since propelled its clinical application have built on the physical principles underlying the magnetic moment and its detection used in the Bloch and Purcell demonstrations.

A key factor in the clinical use of the nuclear magnetic moment is the quantum mechanical effect present in hydrogen protons, which are abundant in soft tissue and confer the equivalent of an angular “spin” on the nucleus [4]. In the presence of an externally applied magnetic field, the magnetic moment generated by the spin processes about the axis of the applied field at a rate proportional to the external field and at a fixed angle to the field’s axis, which defines the Larmor frequency. Within the static field, a second, perpendicular magnetic field applied by radiofrequency can excite the processing nuclear moment, changing the energy level and frequency of the magnetic moment with respect to the static field. In practice, this second field is applied in pulses typically lasting on the order of microseconds. Energy input from the pulses is then re-emitted between pulses as the magnetic moment returns to equilibrium, a process termed free induction decay. Relaxation to thermal equilibrium can occur either longitudinally, with loss of energy to the surrounding environment (a reduction in signal intensity, termed T₁), or transversely, with energy exchange (but not loss) with neighboring nuclei (seen in a broadening of signal phase, termed T₂). T₁ and T₂ values are defined as the lengths of time taken for their respective signals to decline by 63% (or to 37% from the equilibrium value), which adopt first order, exponential decay.

1.2 Spatial resolution and localization

Although the measured signal has its origin in the physical features inherent to the tissue, its detection and subsequent conversion to an anatomical image relies on technical features that amplify, resolve, and reconstruct the unique spatial and physical properties of the image [5]. Localization of the tissue signal, notably, requires the use of magnetic field gradients, which is achieved by varying the static field strength during signal retrieval. In a typical application, the gradient introduces a linear variation in the static field strength along a tissue axis. Because spin precession frequency is directly proportional to field strength, there is a one-to-one correspondence between frequency and position. Gradients can assume any orthogonal direction and, in principle, can be used to reconstruct 3-dimensional images.

Faster or slower precession, and so higher and lower frequency, is detected as a higher or lower signal. Accordingly, manipulating the magnitude of the field strength range modulates the range of frequencies emitted for signal retrieval along a single axis. Imaging data acquired with small gradients has a small frequency range, limiting the frequency differential that can cover a pixel and so also limiting resolution. Larger gradients expand the frequency range, increasing resolution. When plotted, low frequencies are thus distributed centrally and high frequencies peripherally in a mathematical domain termed K-space [6, 7]; all image points thereby contain information from all frequencies. Planar localization requires encoding information along a second axis, a process also achieved by means of a static field gradient, but in which the phase difference along the second axis is monitored. Because phase differences can only be detected with respect to a reference value, a single scan is required for each phase encoded value. Information from both axes must then be mathematically transformed (by Fourier transform) to yield an x-y, planar image.

Because the typical and early use of the method—termed spin echo—sequentially varied the strength of the static field gradient, an important limitation in the use of MRI imaging has been the scanning time required to produce an adequately resolved image. Scanning time is a function of two factors: the interval between radiofrequency pulses, termed repetition time (TR), and the number of scans required for phase encoding and detection. The product of these yields the total scan time. Employing this protocol, typical scanning times could take on the order of minutes to hours, with higher resolution images needing longer intervals and so corresponding corrections for time related artifacts such as patient movement. Because of such artifacts, scanning time has posed a significant obstacle to image acquisition and quality.

1.3 Evolution in MRI procedures

Much of the early evolution in MRI methods thus sought to achieve a balance between data acquisition procedures that provided for sufficient spatial and temporal resolution and contrast and the time required to achieve a desired image. An example of this balancing is seen in the line reductions first introduced to the basic spin echo protocol [6]. A drawback to this approach, however, was the attendant loss in either resolution or image size due to the loss of frequency information. In the fast spin, echo planar or multi echo approach, for example, scanning time was reduced by taking multiple lines after the radio frequency pulse. A chief disadvantage of this latter approach, however, was the rapid loss in signal strength due to energy transferal during T2 acquisition, permitting only 3–4 lines per RF pulse, with significant deterioration in image quality.

While the objectives of reducing acquisition time and achieving enhanced spatial and temporal resolution have been of perennial interest in MRI development, new developments for addressing these and other objectives have in turn revealed new needs that invited corresponding solutions. The widespread use of parallel imaging [8], for example, was due to its ability to accelerate data acquisition while maintaining high resolution through reductions in phase encoding lines and the use of algorithms that corrected subsampling aliasing. Such rapid acquisition methods led to the proliferation of multi-parametric approaches, which then facilitated quantification studies for MR fingerprinting [9]. These, and extended MR functional analyses, for instance, whole brain resting state fMRI [10], have engendered the further development of big data methods for the accompanying computational requirements, while data processing needs for processes such as feature extraction [11] propelled the use of machine learning to reduce diagnostic load. Indeed, the range of new techniques extends from advances in image acquisition to data processing and inference to computation and data handling, enhanced spatial resolution, and functional analyses. To illustrate the current proliferation, this introductory chapter to New Advances in Magnetic Resonance Imaging presents a select representation of techniques that trace their evolution from the physical and detection principles first identified by Bloch and Purcell. Chapters that follow each present a single, recently evolved MRI technique in greater detail. It is hoped that the novelty of these advances will be a source of inspiration for the engaged professional and interested scientist alike.

2. Acquisition procedures: advances in parallel imaging

Parallel imaging is now used in nearly every clinical MRI scan for rapid data acquisition, for numerous reasons [8]. Many abdominal and cardiac scans, for example, are taken with patients holding their breath, with its obvious need for short scanning times. In other cases, for instance, multiline sequences following excitation pulses (e.g., in turbo spin echo), blurring artifacts are introduced due to the substantial T2 decay occurring during line retrieval. In still other applications, rapid data acquisition is essential because of the need to acquire large data sets [12, 13].

In parallel imaging, scanning time is reduced because the phased array coils yield unique views of the tissue objective, eliminating the scanning time for a significant portion of the region subject to gradient encoding [8, 14]. Due to the rapid decline in sensitivity of each coil element with increasing distance from the coil, this limits data acquisition to a clearly delineated, tissue profile. Individual images are combined to yield a comprehensive image. As a matter of principle, the maximum acceleration factor is related to the number of coils. Since most parallel imaging often employs 4–8 coil arrays and arrays containing 32, or even 128, channels are known (e.g., cardiac imaging), the reduction in scan time can be substantial.

Protocols for parallel imaging are currently classed according to whether aliased pixels are segregated in the imaging domain (SENSitivity Encoding or SENSE) or in K-space (Generalized Autocalibrating Partial Parallel Acquisition or GRAPPA) [14]. Techniques that act on the image domain first reconstruct and then correct, while those that act on the frequency domain first correct and then reconstruct. For those segregating in the imaging domain, prior knowledge of the coil sensitivity profiles enables separation of folded pixels from the undersampled image to recover the full image. Recent improvements have included phase-constrained SENSE. Whereas in conventional SENSE unknown variables are complex value, in phase-constrained SENSE, values are real, reducing the variable number by half. In contrast to the SENSE techniques, the GRAPPA algorithm is a K-space technique, which operates on acquired frequency information that is embedded in K-space. The technique is based on the principle that K-space information is shared between points in K-space due to the variation in multiple static field gradients; hence, missing information can be computationally reconstructed from acquired K-space data. In order to reconstruct information from the missing points, acquired data points must be adjusted by weighting using an autocalibration signal obtained from another region of K-space.

Two recent modifications of the GRAPPA technique have included the 2D CAIPIRINHA and Wave-CAIPI techniques [15]. In 2D CAIPIRINHA, acceleration in the K_y and K_z directions is accompanied by a phase offset along K_z. This generates unique frequency patterns that require less computational resources for resolving aliasing. The Wave-CAIPI technique builds on 2D CAIPIRINHA by adding sinusoidal gradients along the K_y and K_z axes with a 180° shift. These additions amplify the acceleration by a factor of 9 over the 2D CAIPIRINHA technique. Unlike other very fast acquisitions, Wave-CAIPI is not subject to blurring from data gridding or artifacts from distortion due to uneven static magnetic fields.

3. Reconstruction procedures for image analysis

The initial stages of image acquisition, preprocessing, and segmentation prepare the data for extraction of meaningful information. By definition, these steps involve removal of nonmeaningful or noise-based signals. One common source of noise, for example, is due to patient movement. Motion artifacts are corrected by registering sequential images, which can be carried out using available algorithms suited for medical imaging. A current standard for MRI image registration is the InsightToolKit (ITK), which contains a suite of algorithms including such processes as transformations, similarity metrics, and contrast normalization [11].

Recent trends in preprocessing and segmentation (e.g., denoising) have employed machine-based learning applications. Due to the labor involved, another major trend in machine-based learning procedures has been that of feature recognition and classification [16]. The high quantity of imaging data acquired from existing MRI scanning, especially, has made clinical diagnoses based on MRI images increasingly laborious, driving efforts for automated data abstraction and analysis. Computationally, machine learning relies on algorithms derived from neural network structures, which are composed of nodes joined by weighted edges. Inputs to nodes are weighted by a set of parameters and multiplied by transfer functions, for example, sigmoid and hyperbolic tangent functions, which transform weighted inputs. Among the most widely used of these deep neural networks (DNN) are the convolutional neural network (CNN), ResNet, the generative adversarial neural networks (GANs), and the U-nets.

Deep network models used for processing whole images are highly complicated, which significantly amplifies the processing time. For example, a CNN training model can often create millions of parameters in searches for feature classification. This drawback has been addressed by the use of image representations of smaller size that increases the processing efficiency. Most current feature extraction applications utilize wavelet transformation techniques in conjunction with neural network processing for MRI images [17]. The wavelet transform is used to remove extraneous detail and make the image more efficient for network processing; hence, approximated images have denser information content than original images. Convolutional layers then apply an initial filtering to produce an initial feature map, that is, what the network deems as unique features, which is refined by further network processing.

4. Multi-information sourcing

Compared to methods for single parameter data analysis, models using multi-parametric techniques offer the chief advantage of assessing correlations between multiple quantitative parameters of interest [9], with the potential for significantly greater accuracy. Various relevant MRI quantities include, but are not limited to, longitudinal (T1) and transverse (T2) relaxation times as well as retrospective synthesis of conventional MR contrasts, which are monitored in conjunction with novel techniques for rapid data acquisition and computational analysis.

Multi-parametric, analytical techniques currently employ similar strategies [18] with simultaneous sampling of the parameter and K-spaces, in which transient-state data are obtained by varying the acquisition parameters and undersampled K-space snapshots are taken after each excitation. Parametric maps are then computed using a physical model via the Bloch equations. Several procedures that have evolved from this approach include magnetic resonance fingerprinting (MRF) [18] and quantitative transient-state imaging (QTI).

In MRF, modulating the MRI sequence parameters across the time domain yields a time series of weighted MRI images, with each tissue having a unique MRI signal fingerprint. Such fingerprints can be computationally simulated and a dictionary of tissue specific fingerprints built from the simulations. During image reconstruction, putative fingerprints are matched to this dictionary. The fingerprint with the greatest correlation in the dictionary provides the MRI parameters for a given voxel. After all voxels are analyzed, parametric maps are then constructed. Due to its ability to recognize very specific structural elements, MRF has the potential for diagnosis of a wide variety of clinical conditions [19, 20, 21, 22]. Novel techniques like quantitative sequencing enable clinical approaches capable of accelerated quantitative mapping of dynamic physiological processes. For example, such techniques have been used to examine blood flow, with computation made of velocity scalars or vectors that could be employed in cardiac assessments. In one study, computations of scalar velocities were computed in a direction perpendicular to a vessel slice based on multi-parametric T_1, T_2, and proton density recordings [23]. A drawback to such methods is that of their reliance on physical models of the physiological events the quantitative mapping is intended to simulate, with the potential for loss of valuable data. Moreover, the complexity of such models can demand increasing computational resources, significantly extending the data acquisition period.

5. Functional approaches using MRI

Blood oxygenation-dependent contrast, termed BOLD or fMRI imaging [24], was developed to indirectly assess neural activity in the brain by monitoring activity-induced changes in blood oxygenation. BOLD exploits the neurovascular mechanism of hyperemia, whereby localized brain activity recruits increased blood oxygenation within the region of activation. In typically used protocols, the image sequence relies on T2* weighting with scan times under 5 seconds to measure the hemodynamic response function. Since its discovery, use of this procedure has undergone extraordinary growth [13, 24].

The dependence of the BOLD signal on neurovascular mechanisms, however, has meant that fMRI is also constrained by limitations inherent in the hemodynamic response. Chief among these is the much slower response time than the underlying neural processes being measured. Temporal information of spiking events is therefore heavily blurred, requiring the use of mathematical processing, like that of the general linear model or experimental block protocols, to infer event related, signal activity [24]. With processing, temporal resolution in the 100 ms range can be achieved, which is roughly tenfold slower than the neural events being monitored.

5.1 High-strength imaging

Another challenge to fMRI is that of the low signal to noise ratio (SNR), a consequence of data acquisition, and constraints limiting the extent of preprocessing that can be performed. One of the technical improvements now being attempted to improve the SNR is the use of high-strength magnetic fields, which enhances the anatomical specificity of imaging. Whereas most scanning is done using 3 T fields, equipment using 7 T fields is becoming increasingly prevalent. At these higher field strengths, it has been shown that less spatial smoothing is required and neural activity in cases of resting state networks displays higher correlation coefficients, indicating greater spatial resolution [25, 26, 27]. On the other hand, use of higher field strengths has several drawbacks, including longer sampling intervals, inhomogeneous magnetic field properties, and the logarithmic growth in specific absorption rate (SAR) with increasing field strength [28, 29].

5.2 Multimodal studies (w EEG/MEG)

Among the techniques used to overcome the temporal limitations of fMRI have been multimodal approaches, which combine fMRI with such methods as the EEG or MEG. Both EEG and MEG display rapid temporal responses, with the capability for resolving neural events at millisecond scales. Use of these approaches in conjunction with fMRI is thus premised on the greatly improved temporal resolution offered by these procedures. Advanced technologies have now been developed to simultaneously record EEG and fMRI signal, which help to understand the relationship between the spatial and temporal characteristics of physiological signals [30]. Nonetheless, in comparison to fMRI alone, combined approaches have seen limited use. In the case of EEG, spatial resolution is greatly inferior to that of fMRI and MEG approaches that suffer from source localization issues. This means that the experimental design or clinical assessment must clearly attribute the signal sources prior to drawing experimental and/or clinical conclusions in such combined approaches.

5.3 DIANA fMRI

The persistence of interpretive difficulties with multimodal approaches has generated a long-standing interest in the development of alternative methods capable of both high spatial and temporal resolution. One recently developed method has merged the detection of ultra-weak magnetic fields generated by neural electrical activity with the fMRI detection of the hemodynamic response [31]. This approach, termed Direct Imaging of Neuronal Activity for functional MRI (DIANA-fMRI), interleaves K-space lines used for imaging the hemodynamic response with a K-space line that directly measures the ultra-weak magnetic field. Millisecond resolution is achieved using fast, low-angle shot (FLASH); gradient-echo imaging; and short repetition intervals (e.g., 5 ms). When carried out at high field strengths (9.4 Tesla), signal to noise ratios are reported to be in the range of 20 to 1. To date, the technique has been used only on animal models.

While promising, the technique suffers from its own unique set of interpretive uncertainties. For example, electromagnetic effects based on neuronal current models appear to be ruled out as these oppose the direction of the observed DIANA response. Moreover, in humans, increasing the stimulus duration does not lead to correlative signal changes, suggesting that the signal may be confounded by other interactions such as inflow effects and subject motion. These uncertainties suggest that further development of the approach will require improved understanding of the biophysical factors contributing to the DIANA signal.

5.4 Resting state fMRI

One of the most significant fMRI developments is the use of fMRI during rest, coined resting state-fMRI (RS-fMRI). RS-fMRI focuses on spontaneous low-frequency fluctuations (< 0.1 Hz) in the BOLD signal that occur in the absence of task-related activities. The functional significance of these fluctuations was first recognized by Biswal et al. [32] in a study in which subjects were told not to perform any cognitive, language, or motor tasks. After determining the correlation between the BOLD time course of a seed region identified by bilateral finger tapping and that of all other areas in the brain, the authors found that fluctuations in the left somatosensory cortex were highly correlated with homologous areas in the contralateral hemisphere. This observed correlation led to their conclusion that such “resting networks” manifested the functional connectivity of the brain. The observation of spontaneous, synchronous fluctuations occurring between brain regions has since led to studies that have identified as many as 7 to 17 other stable networks [33], with 7 consistently agreed upon.

Because characterization of resting state networks (RSNs) in the human brain relies on the analysis of temporal fluctuations in the blood oxygenation level-dependent (BOLD) signal, the delineation of RSNs has been directly dependent on the ability of fMRI to detect neural activity [6]. The dependence on the BOLD signal means that RS-fMRI shares advantages that accrue to fMRI—the ability to monitor neural activity, albeit indirectly—as well as disadvantages that characterize its use. Chief among these limitations is fMRI’s temporal resolution, which is dependent on the hemodynamic response time [34]. Accordingly, a key factor in the use of RS-fMRI is the measurement of neural activity fluctuations rather than spiking events per se.

Early studies of RSN functional connectivities, like that of Biswal et al. [32, 35], relied on the selection of regions of interest based on investigator preferences. However, while the simplicity and interpretability of the ROI technique make it procedurally facile and a frequently adopted approach, the method relies entirely on user-defined ROIs and so is limited for network discovery by its a priori, selected criteria. Due to this caveat and coupled with the evolution of mathematical models and improved computational capabilities, there has been a paradigm shift from that of imposing initial conditions on the data to that of extracting patterns of brain activity directly from the raw time series. The main example of this approach is independent component analysis [36]. In this approach, the time series signal is assumed to be due to multiple spatiotemporal processes that are statistically independent of each other. By extracting the independent signals, various time courses of specific brain regions can be constructed and grouped into maps representative of their spatial distribution. Another approach to the interpretation of RS-fMRI datasets employs graph theory, where activity sources comprise nodes and connectivity defines the edges that link these nodes [10, 36]. Unlike ICA, which focuses chiefly on the strength of correlation between different domains, graph theory characterizes the features of network topology. The graph theory approach describes the interaction between nodes by means of such graph parameters as average path length, clustering coefficients, node degree, centrality measures, and level of modularity. Graph theory is thus a promising technique for exploring the integration and segregation of networks in the brain. Among the topological features studied, modularity, the assessment of the presence of functionally independent units or modules that compose resting state networks, has increasingly been used to characterize functional adjustments occurring during behavior, network perturbations, or pathologies that affect network function, revealing significant alterations in such pathologies as stroke [37] and psychiatric disease [38, 39].

In principle, inferences of causality from directed functional connectivity determinations can be extended to brain-wide neuronal dynamics. Empirical studies from RS-fMRI, for example, show that RSNs are differentiated on the basis of their metastability and synchrony [40]. These and similar observations have led to models of brain function and behavior that predict that the human brain at rest operates at maximum metastability, that is, in a state of maximal network switching. The demonstration of RSN properties like metastability thus suggests that directed connectivity changes may be used to assess the construction of brain states. The methodological question that arises is that of generating a descriptive approach relating functional neuroimaging data to whole brain dynamics. Recent attempts to address this question have adopted two approaches. The first employs a BOLD, data-driven, computational method that leverages the method of recurrence structure analysis (RSA), a mathematical procedure derived from Poincare’s recurrence theorem [41]. This “recurrent” behavior can be described by a recurrence plot method (RP), which allows a matrix-based visualization of recurrent states. The second approach posits the governance of RSN dynamics by a ground state global attractor. This global ground state is mathematically described as a stable stationary solution representing a point of maximal stability in a landscape of stationary points (nodes) that information flows toward or away from [42]. This theoretical framework has been shown to successfully account for the highly structured dynamics arising from spontaneous brain activity in RSNs [43].

6. Conclusion and future directions

Building on Bloch and Purcell’s method for detecting the nuclear magnetic resonance phenomenon, clinical applications have evolved at virtually all stages along the MRI procedural pipeline. Long intervals for data acquisition that plagued early clinical use have been substantially lessened; algorithms for preprocessing with co-registration and denoising have become routine; machine learning for tissue-specific fingerprinting and feature extraction has been introduced, and functional methods for cardiac and brain diagnosis are reliably and regularly used. Future directions will see the roles of adjunct computational resources continue to grow. MRI has traditionally been a procedure for structural tissue determinations, and the assessing of structural differences between normal and trauma or disease states has typically been performed by clinicians. Aiding them will be new algorithms and enhanced computational power for improved feature extraction and pattern recognition capabilities. Clinical diagnoses will increasingly rely on dynamic and physiological parameters associated with functional performance.

In the nervous system, the drive to consolidate improved temporal resolution with the current abilities for high spatial resolution, for instance, will seek to enlist methods successfully used for high spatial resolution in novel ways that match neuronal activity timescales, such as the interleaving of K-space data coupled with theoretical models that enable direct neural activity monitoring. Noninvasive techniques will also introduce perturbation studies to monitor how these influence brain dynamics. Together, these and other innovations will amplify the already leading clinical role of MRI-based applications.