Modern Magnetic Resonance Imaging Modalities to Advance Neuroimaging in Astronauts

METHODS

Procedure

We acquired subject MRI data using a 3T General Electric Discovery 750 (DV26) MR scanner at the University of Calgary’s Foothills Medical Centre. We collected three structural images for use in this study: a BRAin VOlume imaging scan (BRAVO; TR = 6.68 ms, TE = 2.94 ms, TI = 650 ms, FA = 10°, FOV 25.6 cm, 1 mm isotropic voxels, 3 min 44 s acquisition time), which was used as a manufacturer equivalent of the commonly used T1-weighted MPRAGE, a Dual-Echo Magnetization-Prepared RApid Gradient Echo sequence (DEMPRAGE; TR = 916 ms, TE₁ = 2.404 ms, TE₂ = 5.796 ms, TI = 900 ms, FA = 8°, FOV 25.6 cm, 1 mm isotropic voxels, 4 min 31 s acquisition time), as well as an MP2RAGE (TR = 7.176 ms, TE = 2.12 ms, TI₁ = 8 s, FA₁ = 7°, TI₂ = 2.2 s, FA₂ = 5°, FOV 25.6 cm, 1 mm isotropic voxels, 5 min 39 s acquisition time), a variation of the standard magnetization-prepared rapid gradient echo sequence which utilizes two inversion timepoints (Figs. 1 and 2).

View Figure

• Download as Powerpoint
• Download Figure

View Figure

• Download as Powerpoint
• Download Figure

Tissue segmentation was conducted in a region of interest (ROI) centered around the cerebellar tentorium (Fig. 3). The ROI “tentorium” encompasses the medial occipital cortex and cerebellar tentorium, spanning a rectangular prism from MNI −25, −99, −22 to MNI 25, −44, 18, as used by Burles and colleagues.¹⁵ Then, we manually excluded the cerebellum from this ROI, resulting in the ROI depicted in Fig. 3. The cerebellar tentorium is a dural structure which separates the cerebellum from the cerebral hemispheres in the brain and supports the cerebrum from collapsing onto the cerebellum due to the effects of gravity (Fig. 3). On most T1 images of preflight astronauts as well as the typical research subject, the cerebellar tentorium is often misclassified as GM. The CSF shifts resulting from spaceflight¹⁵ cause the tentorium and cerebral GM to spread apart, as typically the ventral portions of the cerebral cortex rest upon the tentorium. Due to this new space between the tissues, the dural tissue is less likely to be misclassified at postflight timepoints, therefore resulting in incorrect claims of GM losses in astronauts in nearby cortical regions due to spaceflight.¹⁵ We utilized a tentorium ROI that did not include the cerebellum.

View Figure

• Download as Powerpoint
• Download Figure

GM tissue segmentation was conducted manually by two tracers, L.B. and T.J., in the tentorium ROI. Here, we manually segmented the GM in the tentorium ROI in spaceflight-naive subjects, in which typical automated segmentation algorithms often misclassify large portions of the cerebellar tentorium as GM.¹⁵ Manual GM segmentation was done using the commonly employed software ITK-SNAP²⁵ in the BRAVO modality of all subjects by both tracers. Manual segmentation of this region was performed to serve as the comparison with the automated segmentation,²⁶ as manual segmentation has been known throughout the literature as the segmentation method prone to the fewest large errors (i.e., systematic bias errors), although extremely labor intensive.²⁷

After manual segmentation, we then utilized automated segmentation procedures from several software tools (Statistical Parametric Mapping 12 [SPM12], FSL 6.0.6, Freesurfer 7.2.0, and Advanced Normalization Tools [ANTs] 2.4.2) for comparison against manual segmentation. Generally, we utilized default or commonly reported settings with minimal adjustment to characterize naive performance of these algorithms with different imaging modalities.

SPM12 unified segmentation module²⁸ was used to independently segment all image modalities. Sampling distance was set to 1 mm, multiple volumes from the DEMPRAGE and MP2RAGE modalities were included as separate channels.

Images from all modalities were independently brain-extracted using the FSL brain-extraction tool (FSL-BET) with robust brain center estimation. The brain-extracted images were then segmented using FSL’s automated segmentation tool (FSL-FAST) with default settings.²⁷

Due to Freesurfer’s limited capability to process multimodal or multivolume MR images as compared to other software packages, the DEMPRAGE and MP2RAGE four-dimensional images were each combined into three-dimensional volumes for processing. The images from the DEMPRAGE were combined by computing the root mean square value of each voxel across the volumes from each echo.²⁹ The images from the MP2RAGE modality were combined using the methodology provided by Knussman and colleagues, which multiplied the uniform image with the inversion (“second”) image to remove background noise.³⁰ The unmodified BRAVO and these processed images were then segmented using Freesurfer’s recon-all command, with no additional parameters. The cortical ribbon images output by recon-all were used as native-space GM segmentations for comparison with other methods. In contrast with the other automated segmentation procedures, the GM output selected from Freesurfer does not include GM from the cerebellum.

In the ANTs pipeline, the first step involved bias-field correcting images using N4BiasFieldCorrection;³¹ each volume in the multivolume acquisitions was bias-field-corrected independently. We utilized the antsBrainExtraction.sh script with templates from the OASIS dataset³² to generate brain masks for each imaging modality for each subject. These masks were then used to constrain segmentation using Atropos, performing a 3-tissue k-means classification.

Prior to analysis, GM probability maps from each MR modality from each segmentation software were binarized using a 0.5 threshold to classify all brain tissue as either GM (1) or nongray matter (0). This results in outputs that mimic the format of the manual segmentation and facilitates their comparison.

Statistical Analysis

Dice coefficients were obtained for each subject across all modalities and automated segmentation software. Dice coefficients were also utilized to quantify the agreement between tracers (Fig. 4). The formula for the computation of Dice coefficients is: $$\mathit{\text{Dice}}=\frac{2x\left|A\cap B\right|}{\left|A\right|+\left|B\right|}$$

View Figure

• Download as Powerpoint
• Download Figure

Dice coefficients are the result of the spatial similarity of two segmentations, and this coefficient attempts to determine the amount of similarity between two (or more) sets of segmented regions or volumes; in the context of neuroimaging, the Dice coefficient is often used to quantify the accuracy of image segmentation algorithms, or between two or more tracers.³³ It can be calculated by taking the overlap between the two segmentations divided by the total number of pixels in both images and multiplying this by two, with higher Dice coefficients indicating greater agreement between the segmentations of two or more raters.³³ The total GM volume of the segmentation of the ROI was also obtained from all subjects in all modalities and software to determine the total number of voxels that were classified as GM to determine which automated segmentation software was the most liberal or conservative in their segmentation compared to the manual segmentation (Fig. 5). The manual segmentation of the MPRAGE modality reported an average total GM volume of 31,430 voxels (L.B. M = 30,558, SD = 5015 voxels; T.J. M = 32,301, SD = 5915 voxels). In order to compare interrater dice coefficients for all modalities against one another, we conducted two-tailed paired samples t-tests between each modality and software program. The significance threshold employed in this study was set at ≤0.05.

View Figure

• Download as Powerpoint
• Download Figure

RESULTS

Other studies utilizing Dice coefficients conclude that a Dice coefficient above 0.70 represents adequate agreement among individuals, and hence our tracing agreement (M = 0.84, SD = 0.03, MIN = 0.79, MAX = 0.89) is satisfactory and can represent our “ground truth” measure.³⁴ The average Dice coefficient between both tracers (Fig. 6) for the traditional MPRAGE scan was 0.84 (SD = 0.03). This average Dice coefficient suggests that both raters overlapped in their segmentation approximately 84% of all traced voxels. The MPRAGE modality was analyzed across segmentation software, and after statistical comparison with the manual segmentation, it was observed that the segmentation derived from SPM12 was the most similar to manual segmentation for the MPRAGE modality. SPM12 was followed by FSL, ANTs, and then Freesurfer, which was the most unlike manual segmentation (Table I). The total ROI GM volume in MPRAGE was highest and therefore most liberal in SPM12, followed by ANTs, FSL, average manual segmentation, and finally Freesurfer with the most conservative GM classification (Table II).

View Figure

• Download as Powerpoint
• Download Figure

Table I. Average Dice Coefficients Across Software.

View Fullscreen

Table II. Total Volume.

View Fullscreen

The results of a paired samples t-test indicated that DEMPRAGE outperformed MP2RAGE when using FSL and Freesurfer, however the opposite was true when using SPM12 and ANTs (Table III). The DEMPRAGE modality Dice coefficients were analyzed next, and results indicated that they were significantly less similar to manual segmentation than the MPRAGE modality for SPM12 and FSL. However, DEMPRAGE significantly outperformed MPRAGE in Freesurfer, and nominally outperformed MPRAGE in ANTs as well (Table III). The total GM classification for DEMPRAGE was most liberal again in SPM12 and ANTs, followed by FSL and Freesurfer (Table II).

Table III. Dice Coefficient Comparison.

View Fullscreen

Finally, results indicated that MP2RAGE was less comparable to the manual segmentation in SPM12 and FSL, as well as Freesurfer. However, in ANTs, the MP2RAGE modality was shown to be more similar to manual segmentation than MPRAGE (Table I). Total volume calculations for the MP2RAGE modality were most liberal using ANTs, followed by SPM12, FSL, and Freesurfer (Table II).

DISCUSSION

This study explored the use of DEMPRAGE and MP2RAGE modalities that were hypothesized to perform more similarly to a “ground truth” manual measure of segmentation,^18–20 resolving a newly established segmentation issue present among astronauts following a spaceflight mission.¹⁵ To test this hypothesis, we compared the performance of automated software versus ground-truth manual segmentations for the three different image modalities. Contrary to our expectations, our results did not establish an obvious MR imaging modality that outperformed across all automated software. Instead, our findings suggest that the traditional MPRAGE modality still outperforms DEMPRAGE and MP2RAGE in terms of similarity to the manual segmentation in the automated software SPM12 and FSL. However, our results did indicate that DEMPRAGE and MP2RAGE perform most similarly to manual segmentation when using Freesurfer and ANTs, respectively. These findings are not surprising, as the literature reports challenges when selecting MR modality for automated segmentation, suggesting that no single methodological approach could be suitable for all images, not all methods could be conceived as equally effective for a particular type of image,²⁷ and there is no gold-standard software package to be adopted for brain segmentation.^17,35

In order to determine which modality resulted in the most liberal and conservative GM classification among segmentation software, we analyzed their total volume estimates. Upon examining the results, a pattern emerged indicating that Freesurfer was extremely conservative in GM classification compared to all other segmentation software. One potential explanation for this bias is that although the cerebellum was removed in all ROIs, some cerebellar voxels may have possibly eluded this deletion. Freesurfer may have avoided segmentation of the cerebellum more accurately than others, resulting in less voxels being classified. Hence, it is a possibility that the Dice coefficients were also affected by this conservative bias in GM volume, as on average Freesurfer had significantly lower average Dice coefficients than other software in MPRAGE and MP2RAGE, and only 0.002 from the lowest Dice coefficient for DEMPRAGE. This finding seems to suggest that Freesurfer may have reported less GM on average than other software, or simply was most conservative in its GM classification, resulting in a lower similarity as less GM voxels were reported. Various literature echoes this concern, as some researchers report that Freesurfer underperforms in regard to robustness and consistency of automated segmentation in comparison to other software.^36,37

Another possible explanation for our results indicating that these newer modalities only perform better in certain software may be that the automated segmentation software we selected had not been developed with the purpose of segmenting these newer imaging modalities.¹⁷ These software were, presumably, developed specifically to segment BRAVO/MPRAGE MR images.³⁸ Hence, it is possible that the default segmentation parameters used in traditional MPRAGE modality are not optimized for the newer scans and may consequently impact segmentation performance.¹⁵

This concept was further supported by visually inspecting the newer modalities (Figs. 1 & 2). In Figs. 1 and 2, it appears significantly easier to differentiate the tentorium from the surrounding GM in the newer DEMPRAGE and MP2RAGE modalities than in the traditional modality. This visual observation suggests that image sequence parameter optimization is possible and may improve the accuracy of these newer modalities. The images obtained using default parameters on the GE MRI scanner and could potentially be improved.

As previous research has suggested that general pipelines were sufficient or were vague in their methods of pipeline optimization,^10,20,39 we did not investigate pipeline optimization and utilized default settings for all modalities to provide a baseline measure of segmentation accuracy. Future research could expand upon the default settings to better optimize segmentation performance in automated software. There are possible adjustments that could improve performance of automated segmentation using the newer DEMPRAGE and MP2RAGE modalities. SPM12 has the option of utilizing Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) to improve normalization and segmentation precision, which has been demonstrated to be effective for segmentation research.^10,19,38 Another method to improve segmentation in SPM12 is to alter the number of Gaussians representing each tissue class and empirically determine which is best for each modality. This was recommended by Viviani and colleagues to fully realize the potential of multimodal segmentation, and it will be important to explore models of signal density in which the number of Gaussian components is varied to identify additional features of tissu.²⁰ In FSL, another potential optimization parameter involves the use of priors as dura is in a similar anatomical location irrespective of modality. Our research did not make use of priors in FSL as its default implementation does not use spatial priors. Finally, computing novel derivatives of raw images may also best capture the extra information DEMPRAGE and MP2RAGE provide, such as average and difference images (Fig. 2). Utilizing this extra information may also provide better segmentation results.

Our study was conceived with the intent of improving brain tissue segmentation accuracy in the astronaut population, however, non-astronaut data were utilized in this research. The primary reasoning for collecting data from non-astronaut subjects involved the challenging nature of sampling the astronaut population. Nonetheless, this factor does not theoretically impact the results or even the generalizability of this work, as the cerebellar tentorium region is error-prone in all individuals due to its proximity to surrounding GM tissue, regardless of occupation.

Another limitation of this work involved the analysis of only one ROI (i.e., the tentorium). Future work should analyze additional ROIs, including the cerebral falx. This dural structure has been, in fact, shown to be impacted by spaceflight specifically, so it would be a key ROI to investigate in data obtained in the astronaut population.¹⁵ Since we have demonstrated in a previous study that spaceflight induces CSF shifts in the brain that cause the cerebral falx to crowd neighboring tissue and results in a similar GM/dura segmentation issue as reported here,¹⁵ an in-depth analysis of the falx ROI may shed light on the impact of this CSF redistribution and tissue-crowding on automated segmentation accuracy. While the tentorium ROI used in this study provided a deeper understanding of the impact of CSF redistribution that results in increased space between dura and surrounding tissues, the falx ROI would indicate the effects of CSF shifts resulting in decreased space between tissue types, potentially reflecting different segmentation inaccuracies.

Another noteworthy limitation of the present study involves the limited MR modalities obtained and analyzed. DEMPRAGE and MP2RAGE were chosen based on previous support in the literature.^17,20 However, there are numerous other MR modalities that may improve the accuracy of automated segmentation performance. One example is the fast GM acquisition T1 inversion recovery (FGATIR), which has been shown to acquire images with higher resolution and sharper delineation of brain regions than T1 or T2 imaging.⁴⁰ The drawback of obtaining FGATIR images is the acquisition time, as it takes approximately twice the time to obtain compared to a MPRAGE acquisition time at 3 T.^40,41 Future work should investigate this modality among others that have the potential to resolve these issues possibly without the use of parameter optimization.

Finally, another prospective solution to this issue is to utilize a Gadolinium-based contrast agent (GBCA) during data collection. GBCAs accumulate in the meninges in the outer layer of the blood brain barrier, permitting easier segmentation from GM in T1-weighted images.⁴² There is support in the literature for GBCAs in terms of contrast of blood vessels⁴³ and dura⁴⁴ among other tissues such as tumors. One major concession of GBCA use is that it poses a slight potential risk, as it must be filtered out of the body through the kidneys. Hence, certain populations would be ineligible for this contrast such as those undergoing dialysis, as they rarely have been shown to cause kidney damage.⁴⁵ Due to this complication, it is unlikely that GBCAs will present the universal solution to this segmentation problem.

Space exploration is a flourishing industry, with more humans going into space than ever before. In accordance with this rapid development and expansion, there is an increased obligation to better understand the impact space travel has on the brain. This work contributes to providing recommendations for those investigating this impact through the analysis of two newer MR modalities among manual tissue segmentation. However, the astronaut population is not the sole population that could benefit from this research since other neurological issues, such as traumatic brain injuries (TBIs), result in CSF changes in the brain.⁴⁶ This population is more appreciable, as approximately 1.5 million Americans are diagnosed with a TBI every year.⁴⁷ MRI is a commonly utilized diagnostic tool for neurological conditions such as TBIs, and hence the implications of this research are critical due to their prevalence. The accuracy of MRI segmentations is significant, as they aid in a clearer understanding of the neurological effects of such a common brain injury.

Patients receiving dialysis treatment are another potential population that could benefit from this research, as CSF volumetric decreases are reported in those undergoing dialysis.⁴⁸ This symptom causes similar crowding issues as spaceflight on the cerebral falx, as the space between certain regions decreases, causing segmentation to become increasingly difficult.¹⁵ If individuals receiving dialysis were to sustain a neurological injury, it is very possible that the MRI segmentation would be erroneous. Hence, the research reported in this manuscript has important implications for many individuals and is not limited exclusively to astronauts.

The primary objective of this research was to pinpoint the MR modalities required to understand and solve the segmentation issues present among astronauts. Our findings revealed that advanced MR modalities could be collected to address this issue. However, this could not be the ultimate solution for the segmentation issues detected in the astronaut population, as the scans investigated in our study only improve segmentation in certain automated software. Further investigation needs to be done to determine whether specific parameters or alternative MR modalities could enhance segmentation precision across all major automated software, which would provide a better methodological approach to advance our understanding of the impact of space travel on the human brain.