Cross-validation of a CSF MRI sequence for calculating brain volume by comparison with brain segmentation methods
Lisa A. van der Kleij1, Jeroen de Bresser1, Esben T. Petersen2, Jeroen Hendrikse1, and Jill B. De Vis1

1Department of Radiology, UMC Utrecht, Utrecht, Netherlands, 2Danish Research Centre for Magnetic Resonance, Centre for Functional and Diagnostic Imaging and Research, Copenhagen University Hospital Hvidovre, Hvidovre, Denmark

Synopsis

We recently introduced a CSF MRI sequence to automatically measure intracranial volume (ICV) and brain parenchymal volume (BPV). This sequence with short imaging time (0:57 min) and fast post processing correlates well with qualitative brain atrophy scores. This study demonstrates that the low resolution and high resolution CSF MRI sequences perform well in the assessment of BPV and ICV, with a precision similar to the conventional brain segmentation methods FSL, Freesurfer and SPM. The CSF MRI sequence showed a good to very good correlation with the conventional segmentation methods for ICV and BPV.

Background

We recently introduced a cerebrospinal fluid (CSF) MRI sequence to automatically measure intracranial volume (ICV) and brain parenchymal volume (BPV)1. This sequence with short imaging time (0:57 min) and fast post processing was shown to correlate well with qualitative brain atrophy scores, which suggests that it could be used in clinical practice as a fast quantitative approach to determine brain atrophy. However, the exact performance of the CSF MRI sequence should still be compared to the conventional quantitative segmentation methods. Therefore, the purpose of this study was to validate the CSF MRI sequence in quantifying intracranial volume (ICV) and brain parenchymal volume (BPV). First, the precision of two implementations of the CSF MRI sequence (low and high resolution) was determined. Subsequently, a cross-validation of the CSF MRI sequence was performed by comparison of the measurements with other commonly used brain segmentation methods.

Methods

For this study ten healthy volunteers (2 females) with a median age of 28 (range 24-41) were included. The scan protocol (3T MRI, Philips) was repeated twice, in between both scan sessions the participants were repositioned. The scan protocol included a 3D T1-weighted sequence (matrix 256x256, FOV 232x256x192, TR = 8187ms, TE= 4.5ms, scan time 6:47 min) and a low and high resolution CSF MRI sequence.

The CSF MRI sequence relies on the long transverse relaxation rate of the CSF (T2,csf) to obtain volumetric CSF measurements which are then used to measure ICV and calculate BPV. To enable this, an MLEV T2 preparation scheme is used which allows for the calculation of the T2 in each voxel by measuring the T2 decay (figure 1). For this a τCPMG of 70ms and 0, 4, 8 and 16 refocusing pulses creating effective TEs of 0, 280, 560, and 1120 are used. Next, by comparing the signal in a voxel to the signal in a pure ventricular CSF voxel, the partial volume of CSF within each voxel can be estimated from which the CSF volume can be derived. The scan parameters of the low resolution CSF MRI sequence were: voxel size 3x3x7 mm3, TR 5208 ms, scan time 0:57 min1,2; and of the high resolution CSF MRI sequence: voxel size 1x1x3.5 mm3, TR 6266 ms, scan time 3:21 min.

The 3D T1-weighted images were used for brain segmentation by Freesurfer(5.3.0), SPM(12) and FSL(5.0). In FSL, the f-parameter for the brain extraction tool was set to 0.2. For Freesurfer and SPM standard parameter settings were used.

Statistical analysis was carried out with R (R Foundation for Statistical Computing, Vienna, Austria).

Results

There was no significant difference between the brain volume of the first and second scan for every method (Wilcoxon signed-rank). Whereas the mean brain volumes with the CSF LR, FSL, SPM and Freesurfer were in the range of 1164 – 1265 cc, the highest mean brain volume obtained with the CSF HR was 1430 cc (table 1).

The precision of the CSF HR was better than the precision of the CSF LR for ICV and BPV (table 1). The precision for the ICV (CSF LR and HR) and BPV (CSF HR) was in the range of the segmentation methods. The precision of the CSF LR for the BPV was outside the range of the segmentation methods with a mean absolute difference of about 10-15 cc higher than the other methods. The lower precision of the CSF LR can be explained by two outliers; see subjects 1 and 5 in figure 2. However, no deviating results for ICV were found in these subjects (figure 3). No motion was detected upon visual inspection in these two subjects. Both the CSF LR and CSF HR measurements showed a good to very good correlation with the other segmentation methods for ICV and BPV (see table 2). Overall, the HR CSF sequence correlated better than the LR CSF measure with SPM, Freesurfer and FSL. This difference can again be explained by the two outliers of the LR CSF sequence. The correlation between the CSF MRI sequences and the segmentation methods was overall lower for ICV than for BPV.

Conclusion

This study demonstrates that the LR and HR CSF MRI sequences perform well in the assessment of BPV and ICV with a precision similar to conventional brain segmentation methods. The CSF MRI sequence provides quantitative BPV and ICV measurements in a very short imaging time of less than a minute

Acknowledgements

No acknowledgement found.

References

1. De Vis JB; EurRadiol 2015.

2. Qin Q; MagnReson Med 2011.

Figures

Table 1

CSF LR = CSF low resolution MRI scan; CSF HR = CSF high resolution MRI scan


Table 2

CSF LR = CSF low resolution MRI scan; CSF HR = CSF high resolution MRI scan


Figure 1 Postprocessed images. One slice from subject 8 from A. CSF LR B. CSF HR C. FSL SIENAX

Figure 2 Brain parenchymal volume (BPV) with each method for both scans per subject. CSF LR = CSF low resolution MRI scan; CSF HR = CSF high resolution MRI scan.

Figure 3 Intracranial volume (ICV) with each method for both scans per subject. CSF LR = CSF low resolution MRI scan; CSF HR = CSF high resolution MRI scan.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1164