Automated segmentation of cerebral deep gray matter from MRI scans: effect of field strength on sensitivity and reliability

Fourteen patients with MS and 15 normal controls (NCs) were recruited to undergo brain MRI at both 1.5T and 3T. Table 1 shows the demographic and clinical characteristics of the subjects. The two groups differed on age (p = 0.0009) but not sex (p = 0.43). Patients met the International Panel criteria for either MS or a clinically isolated demyelinating syndrome (CIS) [16]. All patients underwent an examination by an MS specialist neurologist including evaluation of the Expanded Disability Status Scale (EDSS) [17] score and timed 25-ft walk (T25FW) [18]. At the time of MRI, 10 patients were on disease modifying therapy (DMT), while 4 patients were not. Among the DMTs used, four subjects were on dimethyl fumarate, three were on natalizumab, and one each were on fingolimod, glatiramer acetate, or cyclophosphamide. None of the DMTs were started in the 3 months before MRI. Our hospital’s human research ethics board (The Partners Human Research Committee) approved this study and written informed consent was obtained on all subjects. This work was presented in preliminary form at the 2015 annual meeting of the European Committee on Treatment and Research in Multiple Sclerosis (ECTRIMS), Barcelona, Spain; and at the 2016 annual meeting of the American Academy of Neurology, Vancouver, Canada.

All subjects underwent brain MRI at 1.5T (Signa; General Electric, Milwaukee, WI) and 3T (Skyra; Siemens, Erlangen, Germany). Scanner and acquisition details are show in Table 2. On both platforms, we obtained high-resolution 3D T1-weighted sequences covering the whole head. These were matched as closely as possible on voxel size and acquisition time, considering practical scan time limits for patient tolerability. Each sequence was optimized for signal-to-noise based on previous clinically-routine development. Each subject had a scan followed by a re-scan on the same day on each platform. Thus, at each field strength, two scans were acquired from each subject, where the subject was removed from the scanner between scans for a few minutes, and was repositioned and rescanned by the MRI technologist. For all subjects, except three, the 1.5T and 3T imaging was performed on the same day for a given subject. For the remaining three subjects, the interval between 1.5T and 3T acquistion was 6, 16, or 47 days. During the study period, there were no intervening scanner upgrades.

All image pre-processing was performed using Jim software (v.7.0, Xinapse Systems Ltd., Northants, UK, http://​www.​xinapse.​com/​). For both the 1.5T and 3T images, the raw sagittal images did not yield adequate segmentation, particularly of the intracranial volume (ICV) cavity (“skull stripping”; data not shown). With optimization work, we determined necessary pre-processing steps, which were the same for the 1.5T and 3T images. First, all original DICOM images were converted to a Neuroimaging Informatics Technology Initiative (NIfTI) format, and their raw sagittal orientation was converted to axial. Then, 170 axial slices were extracted from each scan starting at the first slice showing the top of the head. This provided whole brain coverage in all patients extending to the foramen magnum. DGM (caudate, thalamus, globus pallidus, and putamen) volumes were obtained by a fully automated segmentation pipeline (FSL-FIRST, v. 5.0, The Analysis Group, Oxford, UK, http://​fsl.​fmrib.​ox.​ac.​uk/​fsl/​fslwiki/​FIRST) (Fig. 1). This pipeline was chosen for its free availability, full automation, and utility shown in detecting short term DGM atrophy and treatment effects in patients with MS [5, 6]. The ICV was the sum of gray matter, white matter and cerebrospinal fluid (CSF), which was obtained by applying these images to a fully automated algorithm (SIENAX, v. 5.0, The Analysis Group, Oxford, UK, http://​fsl.​fmrib.​ox.​ac.​uk/​fsl/​fslwiki/​SIENA) [15, 19]. We assessed three volumetric measures of the DGM structures: 1) raw volumes, 2) those that were normalized by dividing by the subject’s ICV (“fractions” [1]); 3) those that were normalized by multiplying the raw volume by the whole brain SIENAX normalization factor (“normalized”).

Statistical analyses included unpaired and paired t-tests, Fisher’s exact tests, and analysis of covariance with age as a covariate. Correlations were reported for MRI associations with age, EDSS score, and T25FW by Spearman coefficients. Sensitivity (differentiation between groups) was reported using mixed model analysis of covariance with age as covariate, comparing the methods for their ability to differentiate MS from NC by the interaction between method and group. Within-subject correlations were compared using the method of Meng et al. (1992) [20]. Reliability was reported using intraclass correlation coefficients (ICCs) [21] with 95% confidence intervals (CIs). The analysis was generated using SAS (v. 9.4, SAS Institute Inc., Cary, NC, http://​www.​sas.​com/).

DGM scan-rescan reliability within groups, comparing field strengths, is shown in Table 3 for all three methods. Reliability was generally very high for both the 1.5T and 3T measurements. Regarding raw volumes, there was perhaps somewhat higher reliability at 3T and, across all comparisons, higher reliability for measuring total DGM than the individual structures. Regarding fractions and normalized volumes, there was a suggestion of higher reliability at 1.5T for the caudate. Comparing the three methods, there was slightly higher reliability for measuring raw volumes than fractions or normalized volumes. We determined that the latter were more reliable at 1.5T most likely due to increased accuracy vs. 3T in whole brain extraction, necessary for determination of the ICV and brain size (Fig. 2).

Using all three methods, DGM volumes were higher at 3T than 1.5T in both patients and controls (Tables 4 and 5). This was seen across all DGM structures examined, and significant for all comparisons (except thalamic fractions in the MS group). The effect sizes were generally larger for this field strength difference for raw volumes than fractions (Tables 4 and 5) or normalized volumes (Table 5). This effect was also reflected in total DGM data. For example, the total raw DGM volume was, on average, 3.9% higher in patients and 5.9% higher in controls (Table 5) at 3T vs. 1.5T (both p < 0.0001, Table 4). As shown in Figs. 3 and 4, we explored the possible causes for the increased DGM volume at 3T by performing expert, manual segmentation of one scan each from a healthy control and a patient with MS. The manual segmentations were then overlaid on the automated segmentation maps. The 3T images provided a more accurate (and larger) contour of DGM structures. This was most likely related to improved contrast at the ventricular CSF-tissue interfaces, and to a lesser extent, the gray-white edges.

A comparison of DGM volume and fraction in MS vs. NC is shown in Table 4. 3T showed slightly better differentiation of raw volumes (thalamus: p = 0.02; caudate: p = 0.10; putamen: p = 0.02; globus pallidus: p = 0.0004; total DGM: p = 0.01) than 1.5T (thalamus: p = 0.05; caudate: p = 0.09; putamen: p = 0.03; globus pallidus: p = 0.0006; total DGM: p = 0.02) for detecting DGM atrophy in MS vs. NC. Regarding the normalized volumes, atrophy was not detected in patients vs. NC in any structures at either field strength (p > 0.05) except for the globus pallidus at 3T (mean 4.4 vs. 5.0 ml, p < 0.05). Thus, thalamic and pallidal atrophy was detected slightly more definitively at 3T. Otherwise, the presence of atrophy in the other DGM structures was similarly detected at both 1.5T and 3T. Overall, these field strength differences in the ability to differentiate MS from NC were small (all p > 0.10). Comparing raw, fractional, and normalized volumes, there was slightly better sensitivity for detecting atrophy in the MS group for the raw volumes.

With the exception of the normalized putamen volume at 1.5T (r = −0.56, p = 0.04), DGM raw or normalized volumes and fractions at 1.5T or 3T did not correlate with age; the correlations were also not different between field strengths (all p > 0.05, data not shown). Regarding the correlation between DGM and EDSS scores, significant negative relationships were found with normalized volumes at 1.5T in the thalamus (r = −0.61, p = 0.02), globus pallidus (r = −0.53, p = 0.0498), and total DGM (r = −0.56, p = 0.04) and at 3T in the putamen (r = −0.56, p = 0.04). No significant correlations were found at either field strength between DGM raw volumes or fractions and EDSS scores (all p > 0.05, data not shown). The correlations were not different between field strengths (all p > 0.05, data not shown). Regarding the correlation between DGM and T25FW, with the exception of total DGM normalized volume (r = −0.55, p = 0.04), none of the comparisons reached significance (data not shown).