Functional Magnetic Resonance Imaging of the amygdala and subregions at 3 Tesla: A scoping review

i. Sequence type

The T2*-weighted two-dimensional (2D) Gradient Echo-Echo Planar Imaging (GE-EPI) sequence was the most common acquisition sequence with 85% of studies (n=164) reporting its use exclusively and a further 6% (n=11) of authors not specifically naming their sequence but reporting parameter selections in keeping with GE-EPI. A total of 11 studies used spiral sequences, three of which also used GE-EPI and combined the data from both sequences. Two studies used Spin Echo-Echo Planar Imaging (SE-EPI), one used dual-echo EPI, one used a gradient spin echo technique and one used the Arterial Spin Labelling (ASL) technique. One study acquired data with both GE-EPI and ASL techniques.

A majority of studies (56%, n=108) focused on acquisition of resting state fMRI (rs-fMRI) data while 35% (n=68) focused on task-based fMRI, where participants are asked to view images or actively engage in a task during the scan. Sixteen studies reported both task and rs-fMRI acquisitions, whilst two studies by the same group reported using task data for intrinsic resting state data analysis [40, 41], an approach that has been previously validated [42].

ii. Spatial resolution

VV ranged widely from 7.3mm³ (pixel size =2.44mm² x slice thickness=3mm) to 100mm³ (pixel size = 25mm² x slice thickness = 4mm). Fourteen of the 192 studies aggregated data from multiple protocols and scanners, only one of which used the same VV across protocols. To facilitate presentation and discussion of our results, we used Olman and Yacoub’s proposal of VV ranges. From 1mm³ to 20mm³ was considered high resolution and greater than 20mm to 50mm³ was considered standard resolution [24]. All studies with VV greater than 50mm³ were considered to be low resolution.

Of the 192 studies reviewed, only 16% (n=30) acquired high resolution data exclusively with a further four combining high and standard resolution data. All data acquired in 66% of studies (n=126) was standard resolution with a further three combining standard and low resolution data. Of the remaining studies, 15% (n=30) reported use of low resolution data exclusively and a further three combined low and standard resolution data. Of the low resolution studies, the data from all but one fell into the range of 50-70.3mm³ with a single study reporting a VV of 100mm³.

Slice thickness or through-plane resolution ranged from 2mm to 5mm with 66% of studies (n=127) exclusively using data with the median slice thickness of 3.5mm or less. The protocol in 74% of studies (n=143) did not use a slice gap technique. Studies not specifically reporting a slice gap were presumed to have none. Of the remaining 49 studies reporting a slice gap, four combined data using both techniques.

Slice gaps, commonly referenced as a percentage of slice thickness, ranged from 0% - 33.4% of slice thickness in 88% of the 49 studies using this technique. Of the remainder, two studies employed a gap of 50% (4mm/2mm, 2mm/1mm), two used a gap of 53.8% (2.6mm/1.4mm), one used a gap of 100% (3mm/3mm) whilst one reported a gap of 125% (3.4mm/4.25mm).

iii. Imaging Plane

Axial or axial oblique acquisition planes were by far the most popular choice, with 97% of studies (n=186) utilising this plane. Of the remaining six studies, four used sagittal and two used coronal or coronal oblique planes. Studies not directly reporting an acquisition plane (n=68) were presumed to have acquired axial or oblique axial slices due to this being the most commonly reported plane [43].

iv. Whole brain coverage

Whole brain coverage was obtained in 96% of studies (n=185), with all but six of these also utilising the axial or oblique axial plane. Of the seven studies acquiring limited brain coverage, only five acquired high resolution data and all used axial or oblique axial planes as well.

v. Sequence acquisition times

There was a wide variation in acquisition times (TA) with the shortest sequence being reported at 2 minutes 44 seconds and the longest single TA at 24 minutes. TA was unreported in 41% of studies (n=79) and, of the remaining 113 studies, 20% (n=23) used TA of less than 6 minutes only, 46% (n=52) used a TA of 6 to 8 minutes only, 19% (n=21) used a TA of more than 8 to 10 minutes only and 12% (n=13) used a TA greater than 10 minutes only. Four studies reported TAs from two different TA groups, due to either the use of different TA for task and rs-fMRI acquisitions or the combination of data from different sites and or MRI systems. Two studies combined four separate 15 minute acquisitions for a total of one hour of data.

vi. Radiofrequency Coils

As expected, most studies used data acquired on a single coil type with the most commonly used being the 32 channel phased array in 22% (n=42) of studies. This was followed by the 8 channel phased array with 18% (n=35) and 12 channel phased array with 11% (n=21). Three studies utilised the newer technology 64 channel head/neck phased array whilst one used the 16 channel phased array, and one each used 20 and 24 channel phased arrays. Three studies combined data acquired on different coil types, two of which used data from both 8 and 32 channel phased arrays whilst one combined data from 12 and 16 channel phased arrays. The remaining 9% (n=17) reported use of birdcage, quadrature, custom-designed or other ambiguously-named coil types. In approximately 35% of studies (n=68) the authors reported no information on RF coil type.

Sequence of choice for BOLD signal at 3T

The T2*-weighted Gradient Recalled Echo (GRE) or Gradient Echo (GE) pulse sequence is inherently sensitive to BOLD-related susceptibility effects and is the mainstay in use at 3T due to the predominance of T2* contrast over T2 contrast [3]. The 2D version of the GE Echo Planar Imaging (GE-EPI) sequence, with its high BOLD sensitivity, is considered the gold standard [43]. In addition, its ease of implementation and wide availability in both research and clinical environments has rendered it the sequence of choice for fMRI of the brain [24] and this was evident in our review with over 90% of studies employing this technique. The primary difference between spiral and GE-EPI sequences is the k-space trajectories; spiral sequences traverse k-space from the centre to the periphery whilst GE-EPI employs a Cartesian trajectory of line-by-line acquisition [67]. Spiral k-space trajectory techniques have been shown to be superior to GE-EPI techniques in recovering signal in areas typically affected by susceptibility-induced field gradients such as frontal and medio-temporal areas, with the difference in magnetic susceptibility values at air/tissue interfaces typically being around 8 parts per million (ppm) [68]. However, the use of these techniques is not widespread, as evidenced in our review, partly due to the requirement for lengthy non-standard offline image reconstruction which many clinical research sites cannot support [69]. The use of alternative acquisition techniques was also minimal due to their shortcomings. Arterial spin labelling is limited in fMRI due to its relatively poor functional contrast and temporal resolution, despite its high spatial accuracy [70]. Gradient spin echo techniques are also not commonly used due to their lower sensitivity, reduced brain coverage and greater heat deposition into the patient reported as Specific Absorption Rate (SAR) in watts per kilogram [43].

Data quality – SNR, CNR and spatial resolution considerations

Both SNR and CNR are used as measures in assessment of overall data quality in fMRI. As data provides information about signal fluctuations and stability over a time course SNR is referenced as temporal SNR (tSNR), or mean signal intensity over time, [25, 71]. As previously outlined, SNR and VV are directly proportional [19]. In theory, the use of smaller VV to achieve higher resolution fMRI data results in reduced tSNR per individual voxel which seems intuitively undesirable. Therefore in the compromise between tSNR and achievable spatial resolution when selecting VV, often spatial resolution suffers.

However, there are several other issues to consider in relation to overall data quality when choosing VV. Of primary concern is the location of the amygdala itself. tSNR is diminished by localised main magnetic field (B0) inhomogeneity which results in signal loss and image distortion at air/tissue interfaces. Intravoxel dephasing or signal drop-out occurs when B0 is inhomogeneous within an individual voxel, effectively reducing its SNR contribution [72]. Areas particularly impacted are the inferior temporal regions and sphenoid sinuses adjacent to the amygdala where protons are effectively spinning at different frequencies [73]. Also, with Cartesian k-space trajectories such as that used in GE-EPI sequences, distortion results from voxels being mismapped due to phase errors. In these scenarios, increasing spatial resolution by reducing VV has been shown to improve data quality [21] and potential increases in BOLD signal can be obtained by acquiring smaller voxels in areas such as air/bone interfaces that are traditionally difficult to shim well [19]. Similarly, Olman and colleagues reported that although tSNR was reduced when using smaller VV there was less signal variability across the Medial Temporal Lobe, resulting in more consistent BOLD sensitivity across the region of interest [72]. Further, Robinson et al., focusing their measurements on the amygdala, showed that the SNR penalty incurred in halving the VV by reducing the slice thickness to 2mm was offset by a notable decrease in intravoxel dephasing [74].

Partial volume averaging occurs with larger voxels that encompass a mix of tissue types and its effects can be reduced by matching the VV or spatial resolution to the size of the potential activation during the acquisition in order to maximise functional CNR [75]. Partial volume averaging effects have also been shown to be particularly significant in imaging the amygdala, not only due to its small size but its location adjacent to the Basal Vein of Rosenthal (BVR) which has been shown to give rise to confounding vascular signals when it is partially contained when using larger voxels [32].

As BOLD signal changes are effectively the signal of interest, CNR has been suggested as the primary indicator of data quality [20]. CNR is maximised by reducing noise contributions which have a greater negative impact in fMRI. Unlike background noise in MRI images, noise in fMRI time series data is a blend of physiological noise such as respiratory and cardiac pulsation and thermal noise being emitted by the MR system components as well as from the patient’s body [20, 76]. Maximal CNR has been shown to be achievable with higher spatial resolution imaging in conjunction with mitigation of noise contributed by artifactual and physiological signal fluctuations [19]. Physiological noise contributions also reduce tSNR levels and can be addressed during data post-processing and analysis [77].

Taken together, these technical points provide a solid basis for utilising smaller VV in fMRI focusing on the amygdala and it was somewhat surprising to find that only 16% of studies in this review had utilised high resolution data in reporting their findings.

Selection of slice orientation and acquisition method

Data acquisition can be performed sequentially or in an interleaved fashion. Researchers typically use an interleaved slice acquisition technique which is advantageous as is allows the use of a shorter repetition time (TR) and precludes the use of a slice gap [78]. Using a sequential slice acquisition method without a slice gap can result in slice excitation leakage or adjacent slice saturation leading to reduced tSNR [79]; however, MR system manufacturers have developed and implemented methods to mitigate these effects [80]. Of the studies reporting use of a slice gap technique, it is unknown whether an interleaved or sequential slice acquisition technique was used as this data was not extracted for the current review.

A general disadvantage of using a gap technique is that spin-history artifacts can occur with through-plane head motion which is more severe with increasing gap size. On the other hand, the gap technique can be advantageous as it allows increased brain coverage whilst keeping within a defined TR and slice envelope. However, when imaging the amygdala, use of a gap technique could result in missing potential activation occurring in tissue in between the imaged slices [81-83]. As the almond-shaped amygdala is very small at only around 10mm in width [32] with an average volume of approximately 1240mm³ [13], the acquisition of interleaved contiguous axial 3mm slices would likely result in amygdala tissue appearing in only 3-4 slices, and even fewer if thicker slices or a slice gap technique was used. This highlights the importance of considering the size of the structure of interest when making protocol decisions about spatial resolution and slice acquisition techniques. Of the studies in this review reporting on the subregions 20% used some or all data acquired with slice gaps, two of which used standard resolution data with 3mm slices and a 33% gap whilst another used combined standard and low resolution data with 5mm slices and a 25% gap.

Slice orientation has the potential to contribute to signal loss in T2*-weighted imaging and selection of acquisition plane should be also considered in relation to the particular brain regions under investigation. The plane introduced by Talairach and commonly referred to as the AC-PC (Anterior Commissure – Posterior Commissure) line is reportedly the most widely used and has been broadly adopted as a standard reference angulation in fMRI acquisitions [43, 84] and this was evident in our review with 97% of studies reporting its use. Acquisition of whole brain data parallel to the AC-PC line is more time-efficient as it requires around 30% less slices compared to the coronal plane [74, 84]. The use of axial slices angled at approximately -30 degrees (that is, with the anterior edge of the slice tilted cranially) has been reported to improve BOLD sensitivity in the region of the amygdala as well as decreasing susceptibility-related signal loss; however, this angulation may be suboptimal for imaging other regions when acquiring whole brain data that is a requirement for FC studies [63, 85]. Only one study in our review utilised this slice angulation and they reported subregional findings using standard resolution data. Although the study was based on task activation rather than investigation of FC, whole brain coverage was acquired with this angulation [63]. Selection of imaging plane in relation to orientation of the structure of interest, especially when using rectangular or anisotropic voxels, is also an important consideration. The use of large voxels that are also anisotropic has been shown to provide inaccurate data on volume and shape of small nuclei as well as a reduction in signal sampling accuracy which is based on the largest voxel dimension. Therefore small isotropic voxels are preferred because of their cubic nature [86, 87].

Earlier studies on 1.5T systems have explored the use of other slice orientations for improved imaging of the amygdala. Merboldt et al., found that optimal BOLD imaging for investigation of the amygdala required data acquisition in the coronal plane [88]. Chen and colleagues also reported that, for the amygdala, the largest susceptibility field gradients occur in the superior-inferior (SI) direction and higher BOLD sensitivity was demonstrated in oblique coronal images in conjunction with frequency-encoding in the SI direction [89]. This was in agreement with Bandettini who noted the use of coronal and sagittal slice acquisitions in many studies for the same reason [19]. However, along with the arrival in the mid-2000s of 3T systems with more powerful gradients, came a number of additional technical considerations. Whilst superior gradient performance has many advantages, use of stronger gradients was restricted when using a coronal orientation in conjunction with SI phase encoding. This was due to reports of peripheral nerve stimulation and potential signal aliasing using this configuration, thereby limiting its use [34, 89]. A further consideration following the primary selection of phase encoding direction is the secondary choice, for example, SI or inferior-superior (IS). In areas affected by susceptibility artifacts this choice has been shown to impact on the accuracy of data analysis. Depending on the direction selected, images can be either distorted by distraction or compression; choosing the direction resulting in signal distraction is preferred as signal within regions of interest that have been compressed cannot be recovered during analysis [72]. In our review only four studies used the sagittal plane and two used coronal or coronal oblique planes. None of the six reported subregional findings and all acquired whole brain coverage with standard or low resolution VV, a protocol decision potentially driven by the conflicting need to acquire more slices for high resolution data when scanning in these orientations.

Whole brain coverage versus a targeted approach

To accommodate the need for high spatial resolution data acquisition, multi-resolution approaches have been proposed in which a lower-resolution whole brain dataset is acquired to identify the site of activation, following which a targeted high-resolution dataset is acquired at that site [75]. In cases where the location of the target is known, it can be time-saving to focus only on the desired area of activation rather than acquiring high resolution data over the whole brain. The latter approach necessitates an increased number of slices for coverage resulting in increased TR and reduced temporal resolution. These issues present a conundrum when considering the optimal acquisition strategy for the amygdala as both high spatial and temporal resolution are required. If the purpose of the study is to investigate FC, whole brain coverage is necessary as activation needs to be interpreted within the context of amygdala connections with other spatially remote brain regions, rendering a reduced-coverage approach impractical [34, 90]. Seven studies in our review used a targeted approach with all acquiring task rather than resting-state data. Of these, five acquired high resolution data and two acquired standard resolution data. Only one of the 21 studies reporting on the amygdala subregions used a targeted approach and this group acquired high resolution data [62].

Optimal sequence acquisition time

Choice of TA can involve compromise for a variety of reasons, with participant population being a major consideration. TA can also differ between task and resting state acquisitions with longer TAs of 15-25 minutes being recommended for FC studies with greater reproducibility [91]. However, Van Dijk and colleagues showed that fMRI with TA of around six minutes was satisfactory for demonstrating FC patterns in the brain and that performing more runs reflected only marginally on data reliability [92]. Of the 113 studies in our review that reported their TA, only 15 used a TA of less than six minutes to acquire FC data and 40 studies used a TA of six to eight minutes. Acquiring more data over a longer TA allows the possibility of discarding portions of motion-affected data whilst still retaining enough temporal measurements for analysis. However, as extended scan times can result in patient movement, TAs of six to eight minutes provide a reasonable compromise.

In the absence of acceleration techniques (discussed below) TA is a relatively simple calculation of the product of TR, number of slices and number of volumes acquired. Temporal resolution in fMRI, effectively the time taken to acquire all slices once, (i.e. one TR) is negatively impacted when acquiring higher spatial resolution data as more slices are required for whole brain coverage. Maintaining the same TA could be achieved by a reduction in the number of volumes acquired [79]. However, the very small signal changes of only around 1-5% make BOLD imaging extremely challenging and decreasing both the number of volumes and the VV in order to maintain the TA would significantly reduce tSNR [25]. In fact, one method used to improve tSNR when seeking to image at higher resolution is to acquire a greater number of time points or volumes by scanning for longer, thus reaping the benefits of signal averaging; however, a doubling of scan time only increases SNR by a factor of √2 [93].

The critical nature of the relationship between TA and tSNR in detection of fMRI effect size has been demonstrated by Murphy et al., whose work outlined the value to researchers of prospectively determining the required effect size to a specific p-value. Using this information together with tSNR values they demonstrated that the TA appropriate to each fMRI experiment can be calculated and that an increase in tSNR can significantly reduce TA for a given effect size [25].

Radiofrequency coil technology

The evolution of RF coil technology and in particular, the introduction of phased array coils has positively impacted SNR levels and data quality [93]. This review revealed use of a wide range of coil types, including traditional volume transmit/receive types through to a variety of multi-channel phased array receivers ranging from eight to 64 channels Multi-channel phased arrays have been shown to provide SNR benefits over traditional volume coils, in part due to their improved filling-factor which relates to coil fit with regard to the region of interest [93]. Of the studies in this review reporting their coil type, 84% used a multi-channel phased array for signal reception.

An early demonstration of the superiority of multi-channel coil technology in the anterior Middle Temporal Lobe was reported by Bellgowan et al. in a comparison between higher and standard resolution slices (2mm and 4mm) acquired with both an older style transmit/receive volume coil and a 16 channel array receiver. They reported significant improvements in tSNR and CNR when acquiring high resolution data but noted this was only possible with the use of the multi-channel array coil [94].

Since their introduction, multi-channel coils have been adapted to accommodate increasing numbers of individual elements and 64 channel phased array coils are now available. A less desirable function of coils with very high numbers of elements is a reduction in depth penetration in line with decreasing element size. Higher SNR is evident peripherally due to the close proximity of the arrays to the anatomy whilst signal drop-off may be seen centrally, resulting in signal non-uniformity across the FOV [93]. Correction algorithms to compensate for this are widely available on commercial MRI systems, the use of which is recommended in studies focusing on deeper sub-cortical structures such as the amygdala [95].

The most commonly used coil type in this review was a 32 channel phased array which is intuitively a good compromise between high SNR and adequate depth penetration for improved signal uniformity, an important consideration as the amygdala is a relatively deep brain structure. Interestingly, in a direct comparison between 32 and 64 channel phased arrays, a study by Keil et al., found that the 64 channel phased array outperformed the 32 channel phased array centrally as well as peripherally in SNR terms when acceleration factors of 3 or more were used. This was due to g-factor (noise amplification) improvements attributed to the reduced element size allowing better spatial distinction of the variation in signal intensities and this was also the case when the Simultaneous Multi Slice (SMS) technique was employed [96]. As the 64 channel head/neck array is relatively new technology, its use may become more widespread as sites look to upgrade their head coils in the future.

Phased array coils possess several advantages over earlier coil designs. Along with the capability for imaging larger fields-of-view with inherently higher SNR, they allow spatial localisation of signals. This provides an opportunity to reduce TA by employing acceleration techniques such as parallel imaging and SMS, a technique which effectively increases the sampling rate [93]. In general, differing coil sensitivity profiles in these arrays allow signals to be spatially localised, thus enabling data undersampling with a concomitant reduction in TA. Data can then be fully reconstructed based on individual coil sensitivity profiles. However, the main advantage of parallel imaging in fMRI is not image acceleration but improved data quality resulting from reduced signal dephasing due to shorter achievable echo train lengths, inter-echo spacing and optimal echo time (TE) selection. The primary disadvantage of its use is the accompanying SNR penalty as a result of undersampling [79, 97]. Even so, Schmidt and colleagues recommended its use in the Medial Temporal Lobes, prioritising the marked decrease in image distortion in this location over the potential loss of SNR and BOLD sensitivity [98].

Conversely, the primary advantage of the SMS technique is a reduction in temporal resolution. In this case, the variations in coil sensitivity profiles enable spatial differentiation of multiple slices which can be excited simultaneously. This allows the use of shorter TRs, meaning that either TA can be reduced or statistical significance improved by acquiring more data points [99]. Whilst the SMS technique does not attract a SNR penalty, its implementation may result in higher SAR levels [78].