The document below reproduces poster #86 as presented at the Fifth International Conference of Functional Mapping of the Human Brain (HBM99) in Dusseldorf, Germany, June 22-26, 1999.
Reference: Neuroimage 9, Number 6, 1999, Part 2 of 2 Parts, p S86.
 

Semi-automated stripping of T1 MRI volumes: I. Consensus of intensity- and edge-based methods

Kelly Rehm 1,2, David Shattuck 3, Richard Leahy 3, Kirt Schaper 2, David Rottenberg 1,2
1Department of Radiology, U. Minnesota, Minneapolis, USA, 2PET Imaging Center, VA Medical Center, Minneapolis, USA, 3Signal and Image Processing Institute, U. Southern California, USA

Abstract

Distinguishing cerebral from extracerebral structures is an important step in many magnetic resonance (MR) imaging applications. Automated and semi-automated algorithms for the removal of scalp, skull, vessels and meninges from T1 MRI volumes often require manual intervention to set parameters, remove extraneous tissue and correct exclusions. We present a semi-automated stripping method that incorporates intensity- and edge-based algorithms and reduces manual intervention by automatically setting parameters and producing a consensus between strip masks generated using different strategies.

Introduction

The strength of the intensity-based method is that thresholds are set automatically and have predictable effects. Its weaknesses include susceptibility to MR non-uniformity and exclusion of low-intensity grey matter and cerebral spinal fluid. The edge-based method, Brain Surface Extraction, BSE (1), is tolerant of MR non-uniformity and retains internal cerebral spinal fluid (CSF) and low-intensity grey matter. Its weaknesses include reliance on parameters the combined effects of which can be difficult to predict, and a tendency for morphological operations to result in "notches" in the cortical surface.

Methods I -- Restricting the target region

Consensus stripping requires an initial manual step: the operator creates a coarse mask that excludes the neck, face, and much of the scalp. This is acheived by intersecting irregular regions of interest drawn on average axial, coronal and sagittal slices.
 

Figure 1. Creating the coarse mask.

Methods II -- Intensity-based stripping

The intensity-based algorithm calculates three thresholds from the T1 volume histogram (which is smoothed to make it tractable) and uses them as follows:
  1. Treat the largest connected region with values greater than the grey/white threshold as "white matter".
  2. Discard voxels with values above the white/fat threshold that are not in the so-called white matter (e.g. orbits).
  3. Define large connected regions above the CSF/grey threshold as potential cerebral matter.
The CSF/grey matter threshold is automatically adjusted until at least three large distinct regions exist -- indicating a separation of cerebral and non-cerebral matter. The largest region is defined to be the threshold-based mask.

Figure 2.

Methods III -- Edge-based stripping

The BSE algorithm uses non-linear smoothing, edge detection, and binary morphology procedures to locate the brain surface. One parameter controls non-linear smoothing, a second controls edge extraction. Parameter ranges that yield satisfactory results for scans obtained from a single scanner and scanning protocol can be defined; however, the "best" combination of settings varies by scan.

Our approach is to define the threshold mask volume as a target metric for adjusting parameters in the BSE algorithm. We calculate masks for a small set of parameter combinations and compare their volumes to the target metric, correcting for the inclusion of low-intensity values. The closest match is defined to be the BSE brain mask.

Methods IV -- Consensus

Taking a consensus approach capitalizes on the strengths of the different algorithms and de-emphasizes their weaknesses. The combination (union set) of the threshold and BSE masks is made and then smoothed and expanded by applying a 3D Gaussian filter (FWHM of 4 mm) to the binary mask and thresholding the result; this is the "consensus" mask. See Figure 3 for representative slices from a threshold, BSE and consensus mask applied to a T1 MRI volume.

Figure 3. Threshold, BSE and consensus masks.


 

Results I -- Mask reproducibility

Six repeat T1 scans of a normal subject were used to evaluate the voxels contributed by each method and the reproducibility of strip masks.

Six masks were produced from each scan: the threshold mask, the BSE mask, the combination mask, and Gaussian-smoothed versions of each.

The six repeat T1 scans were co-registered using AIR 3.0 (2), and mask reproducibility (R) in the common space was defined as the ratio of the number of voxels common to all masks divided by the number of voxels included within all six overlapping masks. Reproducibility for all mask classes is reported in Table 1.

 
Table 1.  Mask reproducibility across six repeat scans.
Mask class Threshold BSE Combination
Unsmoothed    R: 0.85
   V: 1220 cc		    R: 0.90
   V: 1389 cc		     R: 0.92
   V: 1424 cc
Smoothed    R: 0.93
   V: 1499 cc		    R: 0.92
   V: 1529 cc		 * R: 0.93
   V: 1564 cc
R: ratio of intersection to union masks
V: volume of intersection mask
*Denotes Consensus mask

Results II -- Variation in coarse masks


Because the first step in the consensus stripping chain is a manual drawing of outlines, we examined the effect that different coarse masks had on the resulting consensus mask. The MRI scans were registered to a template volume using AIR. A coarse mask was drawn for each volume in the common space and two additional masks were computed as the intersection and union of the individual masks. The set of eight masks were transformed back to the native space of each MRI scan and eight consensus masks were computed for each scan. The reproducibility of the coarse masks and resulting consensus masks are reported in Table 2.

The reproducibility of the coarse masks was 0.91, with a disjoint volume (difference in volume between the intersection and union masks) of 179cc.
 
Table 2.  Consensus mask reproducibility
Scan # 1 2 3 4 5 6
Reproducibility (R) 0.98 1.00 0.98 0.98 0.98 0.99
Disjoint Volume (cc) 22.7 0.53 25.9 18.8 28.9 12.1

The results show that moderate variations in the initial coarse mask have negligible effect on the final consensus mask.

Conclusions

BSE masks uniquely contributed CSF, low-value grey matter and blood vessel voxels to the combination mask for each scan. Threshold masks uniquely contributed voxels in the brainstem and inferior cerebellum, inferior portions of the frontal lobe, and cortical notches.

Minimal operator intervention is required to provide a coarse initial mask, and detailed anatomical knowledge is unnecessary. Subsequent parameter setting is fully automated. Creating a consensus mask from the masks created by algorithms employing different strategies increases the reproducibility of stripping and corrects for known defects in individual masks.


This work has been supported by National Institutes of Health grants NS33718 and MH57180.

References

  1. Sandor S, Leahy R. IEEE Transactions on Medical Imaging, 1997, 16(1):41-54.
  2. Woods RP, Grafton ST, et. al. J. Computer Assisted Tomography, 1998, 22(1):139-52.