3 versus 7 Tesla magnetic resonance imaging for parcellations of subcortical brain structures in clinical settings

7 Tesla (7T) magnetic resonance imaging holds great promise for improved visualization of the human brain for clinical purposes. To assess whether 7T is superior regarding localization procedures of small brain structures, we compared manual parcellations of the red nucleus, subthalamic nucleus, substantia nigra, globus pallidus interna and externa. These parcellations were created on a commonly used clinical anisotropic clinical 3T with an optimized isotropic (o)3T and standard 7T scan. The clinical 3T MRI scans did not allow delineation of an anatomically plausible structure due to its limited spatial resolution. o3T and 7T parcellations were directly compared. We found that 7T outperformed the o3T MRI as reflected by higher Dice scores, which were used as a measurement of interrater agreement for manual parcellations on quantitative susceptibility maps. This increase in agreement was associated with higher contrast to noise ratios for smaller structures, but not for the larger globus pallidus segments. Additionally, control-analyses were performed to account for potential biases in manual parcellations by assessing semi-automatic parcellations. These results showed a higher consistency for structure volumes for 7T compared to optimized 3T which illustrates the importance of the use of isotropic voxels for 3D visualization of the surgical target area. Together these results indicate that 7T outperforms c3T as well as o3T given the constraints of a clinical setting.


47
The availability of 7 Tesla (T) Magnetic Resonance Imaging (MRI) scanners has rapidly increased in 48 recent years (Forstmann, Isaacs, and Temel 2017;Keuken et al. 2018;Ladd et al. 2018). The 49 theoretical benefits of anatomical 7T MRI over lower field strengths can be attributed to the increased 50 spatial resolution, contrast-and signal-to-noise ratios (CNR and SNR, respectively), which collectively 51 result in higher quality imaging within feasible time frames [4,5]. Improved visibility of pathological 52 alterations on 7T has been reported in the literature for brain tumors [6], epilepsy [7], multiple sclerosis 53 [8], stroke [9], and neurodegenerative diseases [10]. However, to what extent increased visibility 54 afforded by 7T has the potential to improve clinical outcomes regarding invasive neuro interventions 55 remains unknown.

57
A promising clinical application of 7T MRI is the target visualization of structures for deep brain 58 stimulation (DBS) surgery (Forstmann et al., 2017;Horn, 2019). DBS procedures target structures 59 within the subcortex, which is comprised of a large number of small, iron and calcium-rich nuclei that 60 are located in close proximity to one another [2]. Relatedly, alterations in biometals such as iron in 61 human tissue are commonly observed in pathological processes, for instance, the occurrence of 62 dopaminergic neurodegeneration of the substantia nigra (SN) in Parkinson's disease (PD). Such 63 changes in the chemical composition can cause disease specific structural alterations in shape, volume 64 and location [12][13][14]. Moreover, the neurophysical properties of both physiological and aberrant 65 accumulation of biometals can be exploited to increase the visibility of structural boundaries with both 66 ultra-high field (UHF) MRI and tailored post-processing techniques, such as quantitative susceptibility 67 mapping (QSM) (Deistung et al. 2013;Neumann et al. 2011;Wang et al. 2017).

69
The main DBS targets for PD and dystonic disorders are the globus pallidus interna (GPi) and 70 subthalamic nucleus (STN) [20][21][22][23]. Identification of the STN benefits from visualization of the border 71 of the SN, which has also been targeted for epilepsy (Loddenkemper, et al., 2001) Also the parcellation 72 of the GPi benefits from visualizing the boundary with the external segment of the GP (GPe), the 73 stimulation of which has been shown to modulate functional connectivity in Huntington's disease 74 patients [25]. Additionally, the red nucleus (RN) is often used as a landmark for identification and 75 orientation of the surrounding nuclei [26].

77
Conventional MRI can fail to capture the detailed local neuroanatomy due a weaker field strength.

78
These limitations can be directly translated into a clinical setting with regards to the accuracy of MRI 79 based targeting protocols for DBS implantations. DBS of the STN has been related to a number of 80 psychiatric, cognitive, and emotional disturbances (Temel, Blokland, Steinbusch, & Visser-Vandewalle, 81 2005). Moreover, a fraction of patients will fail to respond to stimulation and or maintain their 82 parkinsonian symptoms, and may require the removal or reimplantation of their DBS leads (Beric et al. 83 2002;Temel et al. 2005). These failures to appropriately respond to neurointervention can partially be 84 3T v 7T MRI for imaging subcortex 4 attributed to suboptimal placement of the DBS lead as a consequence of both inaccurate visualization 85 of the target and reliance on landmark identification [29].

87
Clinical MRI often includes parallel imaging (PI) techniques to reduce acquisition time which is 88 associated with an SNR penalty. This is warranted for both practical reasons, to improve image contrast, 89 as well as clinical reasons, as patients with movement disorders cannot be scanned for extended 90 periods of time. PI reconstructions result in spatially varying noise amplification, which is reflected in 91 the g-factor. However, PI can result in both g-factor penalties and longitudinal magnetization saturation, 92 which can produce anatomically inaccurate and distorted images [2,30]. Thus, anisotropic voxel sizes 93 are commonly employed in order to maintain a higher SNR in-plane, however this negatively affects 94 the creation of accurate three-dimensional representation of DBS nuclei. Alternatively, 7T can provide 95 higher isotropic spatial resolution and SNR within a shorter time period [31].

97
Additionally, DBS surgeries commonly incorporate intra-operative micro electrode recordings and 98 behavioral testing in awake patients to confirm optimal lead placement [32][33][34]. This is a time-99 consuming procedure and distressing for the patient. The higher spatial accuracy that 7T MRI offers 100 could contribute to more accurate surgical targeting and clinical efficacy. Additionally, it can reduce the 101 length of the surgery and the requirement for reimplantation, while ultimately contributing to the 102 abolishment of the need for awake testing during surgery and dramatically improving patient comfort 103 (Forstmann et al. 2017;Lyons et al. 2004).

105
In this work, we investigate the potential of 7T for improved targeting with a quantitative comparison of 106 3T with 7T MRI scans. We acquired two sets of 3T data; one representative of the resolution of clinical 107 3T (c3T) MRI typically used for DBS targeting, as well as an optimized set of 3T (o3T) and a set of 7T 108 data from the same participants. We would like to clarify that we did not attempt to run the same 109 optimized protocol at 3T and 7T, since this would have increased scan times for the o3T further, 110 resulting in a scan sequence which would not have any practical use in a clinical setting. Important 111 differences between the obtained contrasts are the anatomical contrast and the voxel size and shape.

112
Direct quantitative comparisons were drawn from both manual and semi-automated parcellations the 113 GPe, GPi, RN, SN, and STN. Given the iron rich nature of these deep brain structures, and our previous 114 studies indicating that for such structures QSM outperforms T2*-weighted images we used QSM 115 contrasts for parcellations (Alkemade, et al. 2017). Additionally, a semi-automated parcellation 116 approach was employed to parcellate the GP, RN, SN and STN, in order to identify potential biases 117 occurring with manual parcellations and whether accuracy increases with field strength.

120
Participants 121 10 healthy participants (male = 2, female = 8, mean age = 25.9 y, S.D age = 5.8 y), healthy as assessed 122 by self-reports, were scanned at the Spinoza Centre for Neuroimaging in Amsterdam, The Netherlands, 123 on a Philips 7T and 3T Achieva MRI system, with a 32-channel head array coil. The research was

144
= 240 x 188, 140 slices, SENSEPA = 2, TA = 10:08mins. The main difference between the clinical and 145 optimized 3T scans is the voxel size. Two separate scans were collected with o3T, with a total 146 acquisition time of 21 minutes and 46 seconds. We would like to note that we were unable to match the 147 o3T spatial resolution with that of the 7T due to specific absorption rate (SAR) limitations.

161
Maps 162 T1 and T2*-maps for o3T and 7T MRI scans were created using the following procedure: T1-maps were 163 computed using a look-up table [38]. T2*-maps were computed by least-squares fitting of the 164 exponential signal decay over the multi-echo images of the second inversion. For QSM, the 3T data 165 underwent more extensive clipping at the frontal and sinus regions as compared to the 7T MRI data.

166
This was required since the algorithm is sensitive to non-local artefacts, which are more prominent in 167 these regions on o3T MRI scans. For QSM, phase maps were pre-processed using iHARPERELLA 168 (integrated phase unwrapping and background phase removal using the Laplacian) of which the QSM 169 images were computed using LSQR [39,40]. Scans were reoriented and skull information was removed 170 as described previously (Forstmann et al., 2014). The c3T MRI sequence did not allow the calculation 171 of quantitative T1 and T2* maps or QSM images.

174
Manual parcellation 175 Inspection of the c3T scans revealed that despite the high in-plane resolution, which allowed the 176 identification of the structures of interest in the axial plane, we were unable to create a biologically 177 plausible 3-dimensional reconstruction of the structures of interest due to the anisotropic nature of the 178 voxel sizes. We therefore decided not to pursue further analyses of the c3T MRI scans. Multi-echo data 179 was not acquired, and therefore it was not possible to reconstruct QSM images for parcellations.

181
For o3T and 7T images, manual parcellations were performed in individual space using the QSM 182 images for the GPe/i, RN, SN, and STN by two independent trained researchers. Given the level of 183 familiarity of these raters with MRI data, we concluded that blinding for the scan sequence was 184 impossible. T1-maps and/or MP2RAGE images were used for additional anatomical orientation and 185 identification of landmarks such as the ventricles, pons and corpus callosum. T2*-maps were also used 186 where appropriate. See Supporting Information 1 for the approach used for manual parcellations.

187
Raters were blind to each other's parcellations, and inter-rater agreement was determined by the Dice 188 correlation coefficient (see statistical methods).

191
Semi-automated parcellation was performed for the combined GPe/i, RN, SN and STN with FSL's 192 Multimodal Image Segmentation Tool (MIST) (Visser, Douaud, Jenkinson, et al. 2016;Visser, Keuken, 193 Forstmann, et al. 2016). QSM-maps and T1-images were used as input for MIST. MIST output 194 parcellations were compared across field strength (o3T vs 7T), as well as across parcellation method 195 (manual vs. semi-automated) in order to assess for potential biases in manual parcellations such as 196 order or practice effects.

198
3T v 7T MRI for imaging subcortex 7 The o3T brain extracted T1-weighted and QSM maps were co-registered via a multi-step process,

199
where first whole brain T2*-maps were registered to the corresponding T1-weighted images using 200 FLIRT (as implemented in FSL version 6.0.1) with 6 degrees of freedom, nearest neighbor interpolation 201 and mutual information cost function. This transformation was then applied to the QSM-maps, 202 extrapolated form the fifth echo of the T2* sequence, also with 6 degrees of freedom, mutual information 203 cost function and instead a sinc interpolation. The same transforms were applied to the manual 204 parcellations to allow for direct comparisons with MIST outputs. All registrations were visually inspected 205 for misalignments by comparing the following landmarks: ventricles, pons, and corpus callosum.

207
The 7T MP2RAGEME sequence allowed the calculation of all contrasts from a single sequence and 208 thus in the same space, not requiring any registration steps. The MP2RAGE was used as the whole-209 brain anatomical reference image and the fourth echo of the second inversion was used for the T2*

216
Dice coefficients were assessed to determine interrater reliability (Dice, 1945). Dice scores were 217 compared between o3T and 7T images to test the directed hypothesis that 7T images result in higher 218 interrater agreement as compared to o3T images. The Dice coefficient was calculated as follows: Where |mi| is the size of mask i and |m1 ∩ m2| is the size of the conjunct mask of mask 1 and 2. A 221 conjunct mask of a set of masks M only includes voxels included by both raters (Dice, 1945).

223
Volume calculations

224
For manual parcellations, all volume calculations were performed using the conjunct volume of the 225 individual raters, as described previously. Calculations for manual Dice factors were calculated in the 226 space in which the parcellations were performed (Alkemade et al., 2017;Keuken et al., 2013). Masks 227 from the MIST output were compared with manually parcellated conjunction masks resampled to 0.8mm 228 for the 7T data, and the masks that were registered from T2* to T1 for the o3T MRI data.

231
The anatomical distance between the centers of mass of the individual structures in the left and right 232 hemisphere was assessed, providing a measure for changes in the perceived location of the individual 233 structures across field strengths.

234
Distances were calculated as follows:

240
Contrast to noise ratios

241
Contrast-to-noise ratios (CNRs) of the QSM images were calculated to assess differences in visibility 242 of the anatomical structure under investigation. Intensities of non-zero voxels were extracted using the 243 segmentation_statistics implemented in Nighres [47]. The CNR was calculated as follows:

245
SI is the signal inside the mask, represented by the mean value of all the voxels in the conjunct mask.

246
SO is the signal outside the mask, calculated as the mean value of all voxels that directly border the 247 outside of the disjunct mask (all voxels scored inside the mask of a single rater). σ0 is the standard 248 deviation of the set of QSM intensities in these voxels. This approach was adopted to ensure that voxels 249 outside the mask were not part of the separate individual masks.

252
All statistical analyses were conducted within a Bayesian framework ( is performed independently of the others so we assume multiple comparisons are not a confounder in 261 the present study [55][56][57]. We incorporated a within subjects' approach, and for all analyses data was 262 collapsed across hemisphere.

278
We had no hypotheses on the direction of potential changes in volumes or Anatomical distances across 279 field strengths, and therefore conducted two-tailed paired samples t-tests. These analyses provided a 280 single model testing for a difference either way, compared to the null hypothesis. Where appropriate,

281
we calculated the reciprocal to determine the evidence supporting the null-hypothesis.

283
Manual v semi-automated

284
Similarly, when assessing manual and semi-automated parcellations within field strength (manual o3T 285 v MIST 3T and manual 7T v MIST 7T), two-tailed paired samples t-tests were conducted for CNRs, 286 volumes, and Anatomical distances, which we did not expect to differ.

288
3T v 7T MRI for imaging subcortex 10 The Dice score for the MIST output parcellations is comprised of a conjunction mask including only the 289 voxels selected by both the MIST parcellation and the resampled manual conjunction mask. Therefore,

290
Dice scores were not directly tested across parcellation methods.

292
Semi-automated o3T v 7T 293 o3T and 7T MIST parcellation Dice scores and CNRs were compared with a one-tailed paired 294 samples t-test, under the assumption that both Dice scores and CNRs would be higher for 7T than for 295 o3T, indicating that 7T is subject to fewer biases than o3T. Volumes and Anatomical distances were

296
again assessed with two-tailed paired samples t-tests.

307
The MR contrasts are illustrated in figure 1. QSM contrasts obtained from o3T and 7T sequences 308 allowed for manual parcellation of the brain structures under investigation, resulting in biologically 309 plausible 3D reconstructions (see figure 2). As previously mentioned, the c3T images provided excellent 310 in-plane resolution, though did not reasonably allow for anatomically accurate reconstructions due to 311 the anisotropic voxel sizes. Therefore, no formal analyses were pursued for the clinical scans. All results

312
have been averaged across hemisphere, and presented with a margin of error of <0.1%. See table 2 313 for the results of the manual parcellations, and table 3 and figure 3 for MIST parcellations.

367
3T 377 than 3T, which are 12 and 8 times more likely than no differences or higher CNRs at o3T, respectively.

378
For the GPi (BF10 = 0.66), no evidence was found either way and for the GPe (BF10 = 5.43), substantial 379 evidence was found for increased CNR at o3T than 7T, which was 47 times more likely than higher 380 CNRs at 7T.

384
Dice scores were calculated per field strength, per structure with a one-sided paired samples t-test for 385 manual and semi-automated parcellations. For the GPe/i (BF10 = 631.44), we found decisive evidence 386 that 7T Dice scores were higher than o3T, which was 22501 times more likely than no difference, or 387 higher Dice scores at o3T (referred to as the alternative). For the RN (BF10 = 9.15), we found substantial 388 evidence that 7T Dice scores were higher than o3T, which is 83 times more likely than the alternative.

389
For the SN (BF01 = 0.61), we found anecdotal evidence for the alternative, with increased Dice scores 390 at o3T than 7T which is 3 times more likely than the initial hypothesis that 7T Dice scores are higher 391 than o3T. For the STN (BF10 = 1.04), only anecdotal evidence was found for higher Dice scores at 7T 392 than o3T which was 7 times more likely than the alternative (see figure 3).

394
Volumes 395 Two-tailed paired samples t-tests were conducted to assess differences in the volume of manual

539
Interestingly, the MIST SN had higher a CNR for o3T than 7T. These findings suggest that overall, the 540 semi-automated parcellation procedures that were applied do not appear to rely as heavily on CNR as 541 manual parcellations and are therefore not subject to the same biases as manual parcellation. However, 542 the SN and STN at o3T may be an exception. It may be that for smaller nuclei, semi-automated methods 543 using lower field strengths or images with larger voxel sizes rely more on CNR for identification of 544 structural boundaries, whereas higher field strengths or images with submillimeter resolution instead 545 rely on the spatial information.

547
Applications 548 It is important to consider the relevance of these findings in light of neurosurgical applications. Previous 549 work has shown that while the visualization of the STN at 7T shows increased SNR, target localization 550 is not necessarily improved (van Laar et al., 2016;[66,67]. Our findings indicate that optimization of 3T

551
3T v 7T MRI for imaging subcortex 20 MRI scans through the use of isotropic voxels and QSM do indeed allow for more accurate visualization 552 of the surgical target, a result that is within reach for clinical application without the need for the 553 investment in new hardware [68,69]. We would like to note that despite improved anatomical orientation, 554 individual variation in the internal structure of the STN may continue to require awake testing of patients 555 during surgery to obtain the desired clinical effect. Additionally, we have shown that an o3T scan can 556 be obtained in a timeframe that is sensible within clinical practice and can account for age related 557 increases in pathological iron deposition by using multiple and increasing echo times without 558 superseding SAR limitations. This is a particularly important finding given the limitations of both c3T 559 and 7T imaging, which include proneness to increased geometric distortions which reduce spatial 560 accuracy and increase artefacts, B1 field inhomogeneity, power deposition, and altered specific 561 absorption rates [2,70]. Moreover, patient-related contraindications such as metal and or electronic 562 implants, prostheses and foreign bodies, vascular or renal disorders, weight and claustrophobia can 563 limit the potential patient population able to undergo a 7T MRI [3,71,72]. Thus, while our results indicate 564 that 7T is to an extent superior to 3T, o3T could provide a more clinically viable option.

566
Considerations 567 The cohort tested in this study consists of young healthy participants, and it is well known that older 568 participants and PD patients have increased iron content in basal structures (Wang et al., 2016). Since 569 the effects on QSM increase with age and disease, we may underestimate the clinical relevance of 570 these findings (Alkemade et al., 2017). Moreover, the o3T consists of two separate scans, whereas the 571 7T acquisition includes a multi contrast scan obtained within a single session. A multi contrast scan at 572 lower fields would have resulted in an increased scanning time, and therefore be arguably more difficult 573 for scanning with patient populations, especially those with movement disorders. Additionally, a direct 574 comparison between the 3T and 7T data would require co-registration to the same space involving 575 resampling of the data. Since the outcomes of such a comparison could differ substantially dependent 576 on the registration approach chosen, we decided not to perform such analyses.

578
Conclusions 579 Given the limited availability and compatibility restrictions in the patient population of 7T MRI systems 580 for clinical application, our results have merit for more short-term improvement of clinical neuroimaging 581 procedures for surgical purposes. Finally, the use of isotropic voxels is of great importance in these 582 efforts, and we call for caution in the application of anisotropic voxels.