Magnetic Resonance Fingerprinting of the Pediatric Brain

INTRODUCTION

With continuing advancements of MR imaging technologies, quantitative imaging approaches have gained substantial traction in both clinical and research applications. For example, diffusion-weighted imaging, perfusion-weighted imaging, functional MR imaging, and MR have already been widely used to provide great insights into normal brain development and various neurologic disorders in children.^1–4 Quantitative imaging approaches yield findings that are objective and potentially more reproducible when system-related biases are controlled. Despite these potential advantages, qualitative T1-weighted and T2-weighted images remain the most widely used MR images in clinical practice, and clinical interpretations/diagnoses largely rely on qualitative or semiquantitative visual assessments. T1 and T2 relaxation times are fundamental MR imaging–specific properties that are governed by intrinsic tissue composition, microenvironment, temperature, and magnetic field strength. Compared with conventional MR imaging, directly measuring T1 and T2 relaxation times can potentially provide more quantitative and objective assessments of tissue characteristics and pathologic processes.^5,6 However, technical limitations—particularly long acquisition time—make these approaches more vulnerable to motion and prone to system-related instabilities, hampering their wide clinical adoption.

MR fingerprinting (MRF) is a novel imaging framework using fundamentally different data acquisition and postprocessing schemes from those of conventional MR imaging relaxometry approaches. In brief, traditional relaxometry methods acquire multiple datasets where one imaging parameter, such as flip angle (FA), repetition time (TR), and echo time (TE) is varied for each dataset depending on the MR intrinsic parameter of interest. Subsequently, all datasets are combined to derive either proton density, T1 or T2, respectively. There are several limitations associated with these conventional approaches. First, the total data acquisition time can be long because multiple datasets are needed. Second, because all of the acquired datasets are used to compute tissue parameters, subject motion between scans could lead to inaccurate results. Finally, system-related biases such as field inhomogeneities, could further affect the accuracy of the estimated tissue properties. In contrast, MRF uses a set of pseudo-randomized acquisition parameters including FA, TR, and TE, in a single scan, to generate unique signal evolutions depending on specific tissue parameters. Subsequently, tissue properties can be estimated using a template-matching method by comparing the experimentally acquired signal evolutions with a preestablished dictionary voxel-by-voxel.⁷ This approach allows for reliable and accelerated parallel quantitative measurements of multiple tissue properties using highly undersampled data.^7–9 In the following sections, we introduce the technical background and the novel applications of machine learning (ML) to accelerate MRF and review the clinical applications of MRF in pediatric neuroimaging.

TECHNICAL BACKGROUND

Overview of Basic Concepts

In this section, we first provide an overview of data acquisition and postprocessing methods of the MRF technique.

Data acquisition

MRF has been implemented based on several types of MR pulse sequences.^10–14 In this section, we review different MRF sequences, imaging parameters, and 2-dimensional (2D) versus 3D acquisition.

The first MRF acquisition approach pioneered by Case Western Reserve University was based on an inversion-recovery balanced steady-state free precession (IR-bSSFP) sequence.⁷ The main reasons that IR-bSSFP was originally chosen include its high signal-to-noise efficiency and sensitivity to multiple important tissue parameters, including T1, T2, proton density (M0), and off-resonance frequency. However, the presence of magnetic field inhomogeneities represents a major concern for bSSFP sequences, which could potentially introduce banding artifacts in the acquired tissue property maps. The fast imaging with steady-state free precession (FISP) sequence is also a part of the SSFP family and has been previously used for rapid quantification of water diffusion and magnetic transfer.^15,16 Compared with the bSSFP sequences, FISP is less sensitive to magnetic field inhomogeneities and immune to banding artifacts at a cost of reduced signal-to-noise ratio. Recently, significant effort has been made to develop FISP-based MRF techniques with great success of applying them in brain, abdomen, cardiac, and pediatric applications.^10,17–21 Finally, the MRF framework enables the combination of different types of pulse sequences in a single acquisition to extract multiple tissue properties of interest.^12,22

Because hundreds to thousands of MRF signals are needed in order to achieve accurate estimates of tissue properties, it is imperative to use acquisition approaches capable of acquiring images in a clinically acceptable time. To this end, the highly undersampled spiral approach has been used where one spiral interleaf is acquired for one MRF time frame, yielding highly aliased images (Fig. 1).⁷ Although other k-space sampling schemes such as Cartesian, radial, echo planar imaging, and rosette trajectories have also been implemented in the MRF framework,^21,23–26 we focus on spiral trajectory approaches because they provide high scan efficiency and are widely used.

Fig. 1.

MRF data acquisition and pattern matching. Pseudorandomized acquisition parameters, including flip angles and TRs, are used in the acquisition, and each MRF image is highly undersampled and aliased. The tissue properties for each pixel are then extracted by matching the corresponding MRF signal evolution to a predefined dictionary.

Significant progress has also been made to translate the 2D MRF originally proposed by Ma and colleagues⁷ to 3D techniques capable of providing volumetric quantitative imaging.^13,17,27,28 With the addition of a linear slice-encoding gradient, 3D MRF data are typically acquired sequentially through the partition encoding direction. The same acquisition parameters, such as FA pattern and inplane spiral readouts, are repeated for each partition, and a constant waiting time is applied between partitions for longitudinal magnetization recovery. With the extra encoding dimension, fast imaging approaches have been explored to further accelerate along the partition encoding direction. For example, Ma and colleagues¹³ proposed an interleaved sampling pattern to uniformly undersample data along the slice-encoding direction (Fig. 2). Although this approach creates incoherent artifacts, these artifacts can be mitigated using a pattern matching algorithm. Whole-brain coverage (~14 cm volume) with a spatial resolution of 1.2 × 1.2 × 3 mm³ in less than 5 minutes can be achieved. Nevertheless, the achieved spatial resolution may not be sufficient for clinical applications, particularly for pediatric subjects.

Fig. 2.

Interleaved undersampling scheme in sliceencoding direction for 3D MRF (Undersampling factor, 2).

Tissue quantification using pattern matching

One major feature of the MRF technique is the use of pattern matching to extract tissue properties (see Fig. 1). Conventional MR acquisition approaches, where a fixed and identical set of imaging parameters is used for each k-space line, yield identical magnetization evolution time courses for the entirety of k-space. In contrast, MRF uses a pseudorandomized acquisition approach to generate incoherent magnetization time courses, such that magnetization evolution time courses oscillate in a manner defined by variable acquisition parameters and unique tissue properties. An MRF dictionary consists of all possible signal evolutions either using the Bloch equation simulations or then creating an extended phase graph. Subsequently, the MRF-acquired magnetization evolution time courses are then matched, voxel-by-voxel, to an entry of this dictionary using an appropriate pattern recognition algorithm and from which unique tissue properties such as T1, T2, and spin density can be calculated.

Partial volume analysis using magnetic resonance fingerprinting

It is likely that multiple tissue types are contained in most imaging voxels. To this end, partial volume analysis approaches can be applied to model signal evolution and potentially extract additional information.²⁹ Assuming “n” types of tissues in a voxel, this partial volume problem can be described as follows:

$$S_v=\sum _{i=1}^n{w}_i{D}_i,$$

where $S_{\text{v}}$ represents the MRF signal evolution from one voxel, $D_{\text{i}}$ represents signal evolution for a given tissue with known tissue properties, and $w_{\text{i}}$ is the fraction for each individual tissue within this voxel. This partial volume analysis was applied to calculate tissue fraction maps for gray matter, white matter (WM), and cerebrospinal fluid in brain imaging.⁷ More recently, a 3-component model was proposed to model 3 water pools including myelin water, intracellular/extracellular water, and free water. Subsequently, the myelin water fraction (MWF) defined as the percentage of myelin water to the total water can be calculated. MWF has been shown to provide better detection of myelin content during early brain development.¹⁸

Technical Considerations for Pediatric Neuroimaging

As a new quantitative imaging method, the performance and robustness of MRF has been extensively validated in multiple studies across different human organs. Here we provide a brief review of a few studies that could affect pediatric neuroimaging.

Repeatability and reproducibility

Compared with conventional qualitative contrast-weighted imaging, quantitative MR measurements should yield more consistent results independent of scanners, vendors, and imaging sites. However, experimental results along this direction are often lacking, which has been one of the major barriers for MRF clinical adoption. To demonstrate the robustness of MRF for future clinical applications, multiple studies have been conducted to examine the repeatability and reproducibility of techniques across MR scanners, field strengths, and vendor platforms.^30–32 With 10 healthy volunteers who were imaged across 10 3.0 T Siemens scanners at 4 image sites, the 2D FISP-based MRF yielded 3.4% and 8.0% variabilities for T1 and T2 in solid brain tissues, respectively.³⁰ The intrascanner repeatability was 2.0% to 3.1% for T1 and 3.1% to 7.9% for T2. In another study performed on 9 healthy volunteers at 2 imaging centers, Buonincontri and colleagues demonstrated an excellent repeatability (coefficients of variation: 2%–3% for T1, 5%–8% for T2, 3% for M0) and a good reproducibility (coefficients of variation: 3%–8% for T1, 8%–14% for T2, 5% for M0) for 2D MRF across 5 GE scanners (2 at 3.0 T and 3 at 1.5 T).³¹ Using a 3D MRF method, our group has also demonstrated that the quantitative measures derived from MRF showed improved intrascanner and interscanner variability as compared with that of conventional contrast-weighted MR imaging.³² All of these results suggest that quantitative tissue measurements obtained using MRF can serve as good candidates for longitudinal assessments of pathologies and monitoring treatment responses.

Field inhomogeneities

One of the major confounding factors for accurate tissue property mapping is system-dependent B1 inhomogeneities; this could lead to spatially varying signal behaviors and inconsistent image quality between studies. Multiple approaches have thus been developed to improve the performance of MRF quantification in the presence of B1 field inhomogeneities. In combination with a rapid Bloch-Siegert B1 mapping, Chen and colleagues¹⁹ have shown that accurate T1 and T2 mapping can be achieved by using a 2D FISP-based MRF in abdominal imaging. A similar approach has been adopted in 3D MRF to improve performance of volumetric brain imaging.¹³ Alternatively, MRF with simultaneous B1 mapping has also been developed where sensitivity to B1 was increased by adding abrupt changes in the FA pattern during acquisition.³³ In a recent pediatric study, we incorporated multiple B1-insensitive preparation modules and low-power excitation pulses in MRF. The results showed that transmit B1 inhomogeneities can be mitigated in pediatric neuroimaging.¹⁸ Fig. 3 shows the T1 and T2 maps obtained using 2D MRF, and the results are consistent with the reference maps where B1 field correction was performed. Finally, using different setups for receive coils, we have also demonstrated consistent quantitative measurements independent of receiver B1 field inhomogeneities.³²

Fig. 3.

Effect of B1 field inhomogeneity. The B1 map was acquired using the Bloch-Siegert method. Corresponding T1, T2, and MWF maps obtained with and without B1 correction are shown. The difference maps were calculated using the B1-corrected maps as the reference. (Reprint from Chen et al., Neuroimage 2019)

Motion robustness

It is well known that subject motion during data acquisition can lead to substantial image quality degradation. This is particularly challenging for pediatric MR imaging because young children have difficulty keeping still without sedation. Compared with conventional Cartesian MR imaging, most MRF approaches use a non-Cartesian spiral trajectory for in-plane encoding, which has been shown to be less sensitive to motion when compared with Cartesian encoding.⁷ In addition, using the pattern matching algorithm, 2D MRF also exhibits a certain degree of motion tolerance. As shown by Ma and colleagues,⁷ the motion-corrupted time frames behaved like noise during the pattern matching process, and accurate quantification was obtained in spite of severe subject motion. However, further investigation demonstrated that motion tolerance of 2D MRF approaches depended greatly on the magnitude and timing of subject motion. Alternatively, several studies propose using iterative reconstruction methods to retrospectively correct motions in MRF.^34,35 However, iterative approaches are time consuming and difficult to extend to 3D MRF acquisitions. Another major limitation of the 2D MRF-based approach is its inability to correct through-slice motion because proton signals would move in and out of the image slices. Extending 2D to 3D MRF provides an opportunity to correct motion in all directions. By using a 3D spiral projection acquisition scheme, Kurzawski and colleagues³⁶ demonstrated improved motion robustness of 3D MRF for brain imaging. Our group has developed a new approach combining 3D MRF with fat navigators for motion correction.³⁷ A rapid fat navigator sampling with a spatial resolution of 2 × 2 × 3 mm³ and whole brain coverage can be achieved using non-Cartesian spiral GRAPPA in 0.5 sec. Representative T1 and T2 maps obtained with and without motion correction are shown in Fig. 4. Note that this approach does not increase scan time for MRF. To date, most of the studies focusing on improving MRF motion robustness have been performed in adults, and future studies of children are needed to further evaluate pediatric applicability.

Fig. 4.

Representative T1 and T2 maps obtained before and after k-space correction using fat navigator signal. The subject was instructed to move intentionally during the 3D MRF scan.

Integration with Machine Learning

ML has received increasing interest in the MR imaging community. Specifically, ML has been integrated into almost every step of MRF including data acquisition, dictionary generation, and tissue characterization. In this section, we review some ML approaches that could potentially enhance the applications of MRF in pediatric neuroimaging.

Improving postprocessing speed

One important feature of MRF compared with most other quantitative MR imaging methods is the use of pattern matching to extract tissue properties, which can operate robustly in the presence of substantial noise and motion artifacts. However, this approach is relatively slow and requires large memory capacity to store both image datasets and an MRF dictionary. Therefore, one logical area for application of ML in MRF is to accelerate tissue parameter estimations by replacing pattern matching with ML. Cohen and colleagues³⁸ proposed a fully connected convolutional neural network (CNN) capable of accurate tissue quantification that was 300 to 5000 times faster than that of using pattern matching. This method should potentially improve the workflow of MRF in a clinical setting.

Acceleration of image acquisition

Another limitation of pattern matching is that it treats each voxel independently, and hence does not take full advantage of information acquired in MRF. Because of this limitation, hundreds to thousands of MRF time courses are typically required for accurate tissue characterization using pattern matching, which prolongs the overall acquisition. However, valuable information exists in local regions of each signal evolution, and measures from neighboring pixels could enable better tissue characterization. Advanced postprocessing methods extracting additional features embedded in both spatial and temporal domains can potentially improve performance, and therefore reduce acquisition time by decreasing the required number of time courses for tissue parameter estimations. Deep learning is an ideal solution for information retrieval from MRF measurements. In a proof-of-concept study, we developed an advanced CNN model with 2 major modules, a feature extraction module and a UNet module, to accelerate MRF acquisition with improved tissue property mapping (Fig. 5).³⁹ The feature extraction module consists of a few fully connected layers, which is designed to mimic singular value decomposition (SVD) to reduce the dimension of signal evolutions.⁴⁰ Although SVD functions as a singlelayer linear mapping, the proposed feature extraction module provides a multilayer nonlinear mapping from the signal to the extracted features, which can be used to improve the robustness and accuracy of tissue quantification. The second UNet module is used to capture spatial information of the feature map and finally generate the estimated tissue property.⁴¹ Our results on quantitative brain imaging demonstrate that accurate T1 and T2 relaxation time mapping can be achieved with 4 times acceleration in MRF acquisition.³⁹ A similar approach has been applied to enable rapid submillimeter 2D MRF (0.8 mm in-plane resolution) in ~7.5 sec per slice.⁴² An example using the aforementioned approach in a 10-year-old subject is shown in Fig. 6.

Fig. 5.

Schematic drawing of the CNN model with 2 modules for tissue property mapping.

Fig. 6.

T1 and T2 maps obtained from a 10-year-old female pediatric subject using submillimeter 2D MRF with residual channel attention UNet.

Rapid whole-brain tissue property mapping

Compared with traditional relaxometry MR imaging approaches, MRF has significantly improved overall acquisition speed. Nevertheless, the scan time is still relatively long for volumetric imaging, especially for a whole brain coverage as is often preferred in pediatric neuroimaging. In addition, high spatial resolution (1 mm isotropic or less) is often desired for imaging children, which poses grea technical challenges in efficient data sampling. Combining parallel imaging and deep learning, we have developed a novel 3D MRF capable of achieving high-resolution (1 mm³) and whole-brain coverage (18-cm volume) MRF in 7 minutes.²⁷ Preliminary evaluation of this new approach on 6 pediatric subjects has been conducted. Example images from a 5-year-old subject are shown in Fig. 7. The extracted quantitative measures agree well with the results obtained using pattern matching as well as our previous findings from a similar age.¹⁸

Fig. 7.

3D MRF (1 mm³) acquired from a 5-year-old subject in axial, coronal, and sagittal views. About 140 slices were acquired in 6 minutes, and the quantitative maps were obtained using an adult-trained CNN model.

CLINICAL APPLICATIONS

In summary, MRF offers several major advantages in pediatric imaging when compared with traditional MR relaxometry approaches. First, the acquisition time is shorter. Second, MRF is intrinsically less sensitive to motion because of the use of a spiral trajectory acquisition scheme. This advantage is particularly critical for pediatric applications because it can potentially reduce the need for anesthesia.⁴³ Third, simultaneous acquisition of multiple parameters mitigates errors of misregistration among different scans. Fourth, there are excellent reliability and reproducibility of MRF-derived measurements across scanners/platforms/vendors, making MRF a suitable tool for longitudinal assessment of pediatric developmental trajectories or treatment responses. Finally, the new 3D MRF sequence enables whole-brain coverage with isotropic high resolution.²⁷ Therefore, MRF has gained increasing popularity since its inception in 2013. The latest applications of MRF in several fields of pediatric neuroimaging, including early brain development, brain tumors and epilepsy, are reviewed.^18,44–46

Assessment of Brain Myelination

Understanding healthy brain development should facilitate earlier identification of neuropsychiatric disorders, so that appropriate treatments or interventions can be implemented promptly.⁴⁷ Recently, there has been renewed interest in linking dysregulated myelination to the pathophysiologic mechanisms of neurodevelopmental disorders, such as attention deficit hyperactivity disorder (ADHD) and autism spectrum disorder (ASD). It is well known that ADHD and ASD are associated with widespread altered WM integrity.^48,49 A recent genome-wide association meta-analysis of 20,183 ADHD patients and 35,191 controls has identified 12 independent foci; several of these foci are associated with myelination and oligodendrocyte function and suggest a potential role of dysregulated myelination in the underlying pathogenesis of ADHD.^50,51 Likewise, precocious maturation of myelination has been observed in both patients and a mouse model of autism.^52,53 In patients with ASD, altered WM has been shown in young ASD children but may disappear or reverse later in life.⁵² In addition, an accelerated myelination trajectory of the WM has been observed in the frontal brain of neonatal mice but not seen in adult mice with autism.⁵³ Thus, developing a noninvasive, accurate, and robust technique enabling longitudinal quantitative measurements of myelin can help shed light on the pathogenetic mechanisms of these neurodevelopmental disorders.

In a pilot study, we successfully demonstrated the feasibility of using MRF to quantify brain MWF in 28 typically developing children younger than 5 years without sedation.¹⁸ We measured the MWF in multiple WM regions and demonstrated that MWF is almost undetectable at 0 to 6 months old and gradually increases after about 6 months of age (Fig. 8). Depending on the brain regions, there is also a phase of rapid increase in MWF between 6 to 12 months and 6 to 18 months old. These age-dependent and spatially dependent myelin maturation trajectories are consistent with findings reported in the literature using postmortem brain tissues^54,55; this implies that MRF can be a promising quantitative tool for in vivo evaluation of myelination development in pediatric neurodevelopmental disorders.

Fig. 8.

Representative T1, T2, and MWF maps from 5 subjects at different ages. A similar slice location covering the genu and splenium of the corpus callosum was selected. Both T1 and T2 decrease, whereas MWF increases with age. (Reprint from Chen et al., Neuroimage 2019)

Brain Tumors

Pediatric central nervous system tumors are the second most common childhood malignancy and account for the most common cause of cancer-related death in children. A variety of multimodality imaging techniques have been used to image brain tumors, but the results are variable.^56–63 Because of the advantages of multiparametric capabilities and repeatability, the potential clinical efficacy of MRF in characterizing brain tumors has also been studied in pediatric patients.

In a preliminary study of 23 pediatric brain tumors (19 low-grade glioma, 4 high-grade glioma), De Blank and colleagues⁴⁴ used MRF-derived T1 and T2 values to characterize regions of solid tumor, peritumoral WM, and contralateral WM. They demonstrated that T1 and T2 values can differentiate the solid tumor portions from the contralateral normal-appearing WM, as well as high-grade gliomas from low-grade gliomas.

Epilepsy

Epilepsy is one of the most prevalent chronic neurologic disorders, affecting an estimated half million children in the United States.⁶⁴ The direct annual health care cost with epilepsy ranged from approximately 10,000 to 47,000 dollars per patient and were higher in patients with uncontrolled seizures.⁶⁵ Approximately 30% of epilepsy patients are resistant to medications but are potentially curable by surgery.⁶⁶ Accurate detection and delineation of epileptogenic lesions is important during presurgical evaluation of medically intractable epilepsies.

At the time of writing this review, there have been no dedicated studies of MRF in pediatric patients with epilepsy. In a study of 15 patients with medically intractable epilepsy (2 adolescents and 13 adults), Ma and colleagues reported that additional findings, including additional lesions or subtle signal differences between epileptogenic and epileptogenic lesions, were observed in 4 of the 15 patients when 3D high-resolution MRF images were made available to radiologists.⁴⁶ The additional lesions identified by MRF included 1 mild malformation of cortical development, 2 heterotopias, and 1 tuber. Of note, these were adult patients, and studies focusing on pediatric epileptic patients are still lacking.

Mesial temporal lobe epilepsy (MTLE) is the most common form of focal epilepsy in adolescents and young adults, with hippocampal sclerosis (HS) as the top cause of MTLE. The clinical diagnosis of HS is traditionally made by visual comparisons of abnormally elevated T2/FLAIR signal between the hippocampi, assuming that the contralateral hippocampus is healthy. The false-negative rate of this qualitative approach is approximately 15%, including patients with bilateral hippocampal involvement or early HS with minimal T2 signal changes. It is well known that T2 relaxometry can improve detection of early hippocampal damage and lateralization of HS in patients with MTLE.⁶⁷ In a recent study of 2D MRF comparing 33 MTLE patients with HS (32 adults) and 30 healthy controls, Liao and colleagues⁴⁵ demonstrated that MRF-derived T1 and T2 values can distinguish HS from healthy hippocampus and can improve the accuracy of MTLE diagnosis when compared with traditional visual assessment (96.9% vs 69.7%) (Fig. 9).

Fig. 9.

Representative tissue fraction segmentation maps in a patient with unilateral hippocampal sclerosis. The histogram shows the different T1 and T2 distributions between healthy and suspicious hippocampi. (Reprint from Liao et al., Radiology 2018)

These 2 studies suggest that MRF is a promising imaging technique to improve sensitivity and accuracy of identifying epileptic lesions, which might optimize management and guide surgical strategy in patients with epilepsy.

FUTURE DIRECTIONS AND CONCLUSION

MRF is a promising quantitative MR imaging technique that can calculate unbiased and accurate tissue properties. These tissue values can then be used as biomarkers to differentiate normal and abnormal tissues or for longitudinal monitoring. The advantages of a short acquisition time and motion robustness offer potential clinical pediatric applications. However, most current evidence involves relatively small sample sizes and heterogeneous populations, and further investigations are warranted.

KEY POINTS.

• Magnetic resonance fingerprinting (MRF) is a new quantitative MR imaging technique for rapid and simultaneous quantification of multiple tissue properties.

• MRF can reliably and accurately characterize intrinsic tissue properties, such as T1 and T2 relaxation times.

• MRF has many potential applications in children, including evaluation of brain development and differentiation of normal from pathologic tissues.

CLINICS CARE POINTS.

• MRF can measure multiple quantitative MR parameters in a single acquisition.

• Compared to conventional quantitative MRI methods, MRF is advantageous for pediatric patients due to the shorter acquisition time and better motion tolerance.

• 3D MRF can provide whole brain coverage and high-resolution images.

• MRF is a promising tool for measuring myelination development, differentiating low-grade and high-grade brain tumors, and identifying epileptogenic lesions.