Introduction

FreeSurfer has been developed to segment conventional weighted images of the brain. Segmentations can be imperfect, including voxels of CSF and other fluid within tissue ROIs. Very few fluid voxels can skew means of quantitative relaxometry data but do not affect geometric measurements of volumes and thickness that FreeSurfer has been optimised for. We present a simple threshold-based correction of FreeSurfer segmentation that eliminates the need for laborious manual adjustment.

Methods

10 volunteers were recruited following informed consent and ethics board approval. The scanning protocol included a 1x1x1mm³ T₁ MPRAGE structural volumetric sequence and a 1x1x5mm³ Siemens FISP MRF^[1] sequence.

The FreeSurfer recon-all command^[2] was performed on the high resolution MPRAGE. In 4/10 volunteers, extra cranial non-grey matter like cerebrospinal fluid (CSF) was erroneously included within the ROI of grey matter, particularly in immediate proximity to the sagittal sinus. An example is shown in Figure 1. The FreeSurfer output segmentations were registered to the quantitative maps using a rigid-body registration in FSL^[3] and were used to extract data from the quantitative maps. Grey matter segmentations of the quantitative maps are shown in Figure 2.

To correct these segmentations, maximum thresholds are applied based on mean values of CSF in ventricles, also segmented by FreeSurfer. For T₁, the CSF mean across 10 volunteers and therefore the maximum threshold for tissue was 2700ms. For T₂, it was 255ms.

Results and Discussion

T₁ and T₂ histograms of grey matter voxels for one healthy volunteer are shown in Figure 3, along with the location of the removed voxels. All voxels in tissue segmentations exceeding mean relaxation times in CSF were removed; these are physically unrealistic for tissue and more consistent with CSF or cortical vessels. Maximum thresholds were computed for both T₁ and T₂ maps. Significant inclusion of CSF in grey matter occurred in 4/10 healthy volunteers. Biologically unrealistic T₁s up to 3500ms and T₂s up to 1500ms in the grey matter, shown in Figure 3, are not due to partial voluming of CSF within tissue voxels but due to segmentation errors.

Box plots of T₂ means in each ROI for each volunteer prior to thresholding are shown in Figure 4a, with thresholding to remove CSF voxels are shown in Figure 4b. Mean values are computed after eroding once (removing voxels in the ROI that are not surrounded by other voxels in the ROI) for every image slice. T₁ data are shown in Figure 5.

Comparing Figures 4 and 5, ROIs with variance between volunteers that was impacted most by thresholding were the same for T₁ and T₂. The corpus callosum is a small white matter region neighbouring CSF in the ventricles, and the occipital grey matter partially surrounds the sagittal sinus, where segmentation errors have been demonstrated in Figure 3. Here, we apply thresholds to T₁ and T₂ separately to demonstrate applicability to any quantitative mapping method, but since MRF T₁ and T₂ maps are co-registered, masks of voxels exceeding thresholds of T₁ or T₂ could be combined for accuracy.
In smaller ROIs such as occipital grey matter which neighbour CSF in the brain, thresholding has a clear impact: less variation in ROI means between volunteers. The inter-volunteer variance of mean T₂ in occipital grey matter is reduced by 92% from 57ms to 5.37ms after thresholding. Thresholding by the mean of CSF removes up to 1.8% of the voxels in the occipital grey matter T₂ segmentation, however removing these voxels leads to a 22% decrease in mean T₂.

The exact placement of the threshold is somewhat arbitrary but requires a meaningful basis. Removing all voxels with a T₁ or T₂ exceeding the minimum value of the CSF in the ventricles removes up to 40% data. The mean of the CSF is a conservative threshold; we can be certain any discarded voxels are fluid not tissue. Some fluid may still be present, but enough is removed to prevent erroneous variation between subjects that results from CSF and other fluid contamination in ROIs of some volunteers.

Conclusion

Removing voxels from tissue ROIs that exceed the mean relaxation times of CSF in quantitative T₁ and T₂ maps leads to reduced variation in affected ROIs across 10 volunteers. Thresholds are easily computed and can be applied in a batch process. Removing only up to 1.8% voxels in ROIs with segmentation errors leads to a more than 15% change in T₂ average in occipital grey matter for 4/10 volunteers. Where ROIs do not neighbour CSF, there are no voxels removed and therefore no change in mean; all data was processed in one step.

Acknowledgements

The authors would like to thank the volunteers who participated in the study and acknowledge funding from the Imperial NHS Imaging Department and the CRUK Convergence Science Centre. The authors would also like to thank the Bangerter Lally group, Imperial NHS ImRes Group, the Imperial MRI Physics Collective, and Mathias Nittka (Siemens Healthineers, Germany).