Automated brain MRI morphometry, including hippocampal volumetry for Alzheirmer's disease, is increasingly recognized as biomarker. Consequently, a rapidly increasing number of software tools have become available. We tested whether modifications of simple MR protocol parameters typically used in clinical routine systematically bias automated MR-based brain morphometry.

This study was approved by the local ethical committee and included 21 consecutive patients (13 females, mean age 75.8 ± 13.8 years) undergoing clinical brain MRI for workup of cognitive decline. MR imaging was performed on a whole-body 1.5T clinical MR scanner (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany) using a 20-channel head coil. Three different 3D T1-weighted sagittal volumes (see Figure 1) were obtained using the magnetization-prepared rapid gradient-echo (MPRAGE) pulse sequences^1,2 employing the imaging parameter settings described in Figure 2.

Brain segmentation was performed by two different and established analysis tools, the MorphoBox prototype^3,4 and Freesurfer⁵ version 5.3.0 using standard parameters. Estimated volumes of the following brain tissues and structures were analyzed with the R software package (version 3.1.1) (R Core Team, 2014): total intracranial volume (TIV), grey matter (GM), cortical grey matter (cGM), white matter (WM), hippocampus, ventricles, and cerebellum.

In order to assess the presence of a potential systematic bias in the volumetric results, relative volume differences (RVD) between the reference protocol ($$$V_{r}$$$) and each variant ($$$V_{v}$$$) were computed for each structure as: $$RVD(V_{r},V_{v})=100\frac{V_{v}-V_{r}}{(V_{v}+V_{r})/2},$$where $$$RVD(V_{r},V_{v})$$$ is in the range [-200%,200%].

ADNI-2 and LOCAL- protocols were used as the reference protocol ($$$V_{r}$$$) in all subsequent analysis. Relative volume differences were averaged across the subjects. The statistical significance of thee difference from the zero median in relative volumetric differences was tested using the Wilcoxon signed-rank test (Wilcoxon, 1945), as the differences were not expected to be normally distributed. Bonferroni correction (Dunn, 1961) was used to correct for multiple comparisons.It has been shown that the variance of volume differences does not significantly change across different systems (different field strength, different vendors), but systematic offsets in volumes may be present⁶. Therefore, to compare our results to previously reported scan-rescan reproducibility studies^7,8,9,10, absolute relative volume differences (ARVD) between the reference protocol ($$$V_{r}$$$) and each variant ($$$V_{v}$$$) were recomputed for each structure as:$$ARVD(V_{r},V_{v})=100\frac{|V_{v}-V_{r}|}{|V_{v}+V_{r}|/2},$$where $$$ARVD(V_{r},V_{v})$$$ is in the range [0%,200%].

ADNI-2 and LOCAL- protocols were used as the reference protocol ($$$V_{r}$$$) in all subsequent analysis. Absolute relative volume differences were averaged across the subjects.

Effect of image filtering

The comparison of LOCAL- versus LOCAL+ protocols with the FreeSurfer segmentation software revealed a significant change in WM and ventricle volumes. The respective RVDs for WM and ventricle volumes were: -1.82% (p<0.05), and -0.98% (p<0.05). The, median ARVD of the hippocampus volumes was 2.99 ± 7.70 % (see Figure 4A). Using the MorphoBox segmentation tool, significant changes in TIV and WM volumes were observed. The respective RVDs of TIV and WM volumes were: -0.28% (p<0.05), and -4.84% (p<0.05). The median ARVD of the hippocampus volumes, was 3.39 ± 3.91 % (see Figure 4B).

Effect of spatial resolution - 1.2 versus 1.0 mm

The comparison of ADNI2 versus LOCAL- protocols with the FreeSurfer segmentation tool revealed significant changes in TIV, GM, cGM, ventricles, and hippocampus volumes (see Figure 5A), even though they were hard to identify upon visual inspection (see Figure 3). The respective median RVDs were 2.42% (p<0.01), 3.14% (p<0.001), 4.52% (p<0.001), 2.40% (p<0.001), and -3.81% (p<0.05). For hippocampus volumes, the median ARVD was 4.23 ± 5.97 %. Segmentation results obtained with the MorphoBox segmentation tool revealed significant changes in TIV, and WM volumes (see Figure 5B). The respective median RVDs were 1.78% (p<0.001), and 1.60% (p<0.01). For hippocampus volumes, the median ARVD was 4.55 ± 4.68 %. Note that the relative volume differences correspond to fixed offsets in segmentation results and do not represent scan-rescan variability of each protocol.

A simple change of MR parameters, notably spatial resolution, contrast and image filtering, may systematically bias results of automated brain MRI morphometry by up to 4.8%. This is in the same range as early disease-related brain volume alterations, for example in Alzheimer's Disease¹¹. Automated brain segmentation software packages should therefore require strict MR parameter selection or include compensatory algorithms¹² to avoid MR-parameter-related bias of brain morphometry results.