Compared to standard MPRAGE, MP2RAGE acquisitions [1] provide improved grey/white-matter contrast [2] and exhibit greatly reduced B1 bias [1]. These properties should improve brain volumetric measurements based on the MP2RAGE contrast; however, a major difficulty in processing MP2RAGE data is the salt-and-pepper noise in the image background and cavities stemming from the intrinsic division of the two inversion contrasts acquired with the MP2RAGE. This poses a major problem for the skull-stripping stage of most automated-morphometry software packages. We evaluated three different skull-stripping strategies using the MorphoBox prototype [3] and FreeSurfer [4].

Experiments were performed on a clinical 3T MRI scanner (MAGNETOM Skyra, Siemens Healthcare, Germany) equipped with 20-channel head/neck and 32-channel head coil arrays. 12 healthy subjects were imaged after providing informed consent using the following protocols:

(a) 5:12-minute 3T MPRAGE protocol that resembles the protocol used in ADNI-2 [5]: TR/TI/BW/α= 2300ms/900ms/240Hz/px/9°

(b) 8:22-minute 3T M2PRAGE protocol (TR/TI1/TI2/BW/α1/α2 = 5000ms/700ms/2500ms/240Hz/px/4°/5°, 1.00x1.00x1.00mm³) using a three-fold GRAPPA acceleration (iPAT=3)

For each subject, the measurement session consisted of a scan and rescan using MPRAGE and MP2RAGE protocols with both 32-channel and 20-channel coils. Note, the MP2RAGE outputs the 1st and 2nd inversion contrasts as well as the MP2RAGE “uniform” contrast as described in [1].

Subsequently, images were skull-stripped using both the MorphoBox prototype [3] and FreeSurfer [4] (version 5.3.0) by inputting auxiliary images with low background noise. We investigated 3 different types of auxiliary images:

Method 1: Second inversion contrast

Method 2: Product of the second inversion contrast and the uniform MP2RAGE image as in [6].

Method 3: Denoised uniform MP2RAGE image as described in [7]. This strategy requires tuning a regularization parameter (denoising constant), which we hypothesise to be proportional to the noise variance estimated using the pseudo-replica method [8]. We tested a range of multiples of the variance as denoising constant (N*Var, N=1…30).

To quantify skull-stripping performance, we evaluated the across-subject root-mean-square errors (RMSE) between MP2RAGE-based and MPRAGE-based estimates of the total intracranial volume (TIV). We also evaluated the reproducibility of TIV estimation for each contrast by computing the RMSEs between scan and rescan.

Fig. 1 shows sample skull-stripped images using the above methods. Note that FreeSurfer brain masks exclude extra-ventricular CSF, and the TIV is estimated based on the Talairach transform without being explicitly segmented [9]. Visually, it is hard to distinguish between methods 2 and 3, whereas method 1 exhibits obvious errors in both MorphoBox and FreeSurfer results.

Fig. 2 illustrates the RMSE in TIV estimates as a function of the denoising constant using MorphoBox. Method 1 introduced large discrepancies in TIV compared to the MPRAGE acquisitions (4.10%/4.13% with 20-channel and 32-channel coils, respectively). Method 2 proved to be more similar to MPRAGE-based TIV extractions, yielding RMSEs of 0.94%/0.98%. With an optimal choice of the denoising constant, method 3 achieved 0.59%/1.38% difference to MPRAGE acquisitions. Scan-rescan RMSEs were comparable across the three methods: 0.172%/0.309% with method 1, 0.288%/0.127% with method 2, 0.164%/0.181% with method 3, which were not substantially different from the 0.264%/0.218% achieved using MPRAGE.

Using FreeSurfer, method 1 introduced considerable discrepancies in TIV estimates compared to MPRAGE as shown by the respective RMSEs of 9.35%/9.12% with 20-channel and 32-channel coils, respectively (Fig. 3). Method 2 yielded slightly smaller RMSEs of 7.00%/7.64% (Fig. 3), while method 3 achieved 1.89%/3.46% with optimally chosen denoising constants. Scan-rescan RMSEs for methods 1, 2, 3 and MPRAGE were respectively: 0.069%/0.134%, 0.167%/3.21%, 0.870%/1.32%, and 0.115%/0.127%. There is a noticeable decrease in scan-rescan stability of methods 2 and 3 compared to MPRAGE. In order to get a better insight into the large discrepancies between MPRAGE and MP2RAGE TIV estimates with FreeSurfer, we examined the brain masks as represented in Fig. 1. Fig. 4 illustrates the RMSE in brain mask volumes compared to MPRAGE. Both methods 2 and 3 produced considerably smaller discrepancies in brain mask volumes than in TIV and showed scan-rescan stability similar to MPRAGE.

Using MorphoBox, both methods 2 and 3 achieved similar TIV estimation performance compared to MPRAGE and clearly surpassed method 1. With FreeSurfer, method 3 performed considerably better than methods 2 and 1, however yielding rather poor correspondence with MPRAGE in addition to larger scan-rescan variability. When using MorphoBox, our results suggest in practice to estimate TIV from MP2RAGE data using method 2 since it can be applied retrospectively contrary to method 3. However, caution must be exercised when using FreeSurfer-based TIV measures computed from MP2RAGE data using the methods considered here (e.g., for volume normalization purposes). Future work will aim to investigate MP2RAGE-based volumetry for other brain structures.