Intracranial vessel segmentation from time-of-flight MRA using pre-processing of the MIP Z-buffer: accuracy of the ZBS algorithm
Introduction
When presented with a medical image, a radiologist must detect and extract significant information in the presence of image noise. Noise in this setting refers not only to random variations in the image but also to structural noise (i.e., structures within the image which impede the detection of the desired information). An advantage of volumetric imaging such as X-ray computed tomography (CT) or magnetic resonance imaging (MRI) is that the structural noise in the image is greatly reduced (Goodenough, 2000). However, the advantage of increased image conspicuity is offset somewhat by the large number of resulting slices within the volume. For many imaging applications the objects of interest represent only a small portion of the acquired volume. Review of the slice images to extract desired information can be tedious. This is particularly true in MR angiography (MRA) and CT angiography (CTA) of the intracranial vessels, where the vasculature represents only roughly 5% of the intracranial volume. Additionally the tortuous nature of the vessels results in very little vessel information within any given plane of the image. With improvements to MR hardware, improved pulse sequences and the introduction of post processing techniques such as zero-filled interpolation, the size of the MRA volume data have become so large that analysis of the image cross-sections is tedious if not intractable. Similarly, the introduction of 8 and 16 slice helical CT scanners has resulted in very large CTA data sets with minimal in-plane information. Therefore, efficient review of MRA and CTA data relies on algorithms which preferentially render the 3D vasculature data in a form that can be rapidly appreciated. In this paper we focus on MRA images.
The choice of how to render the 3D MRA data is limited by the relatively high background signal present. The high background signal produces very poor contrast images if a densitometric (X-ray like) projection is used and overlap between vascular signal intensities and background intensities inhibits segmentation. The maximum intensity projection (MIP) algorithm is a simple solution to this problem and has proven to be the most popular rendering algorithm for MRA (Rossnick et al., 1986; Laub, 1990; Anderson et al., 1990). In the MIP algorithm rays are cast through the three-dimensional (3D) data set and each ray is assigned the value of the maximum voxel it encounters along its path. The MIP algorithm is very simple and provides high signal-to-noise ratio (SNR) for large, bright vessels. However, the MIP algorithm, by keeping the maximum background signal, causes the average background intensity in the MIP image to be larger than the background intensity in the original image data and thus small vessels, with low initial SNR, have reduced conspicuity (Brown and Riederer, 1992). Additionally, since only the maximum value is recorded, the MIP images provide no information regarding vessel overlap or depth of the vessels.
If the vasculature was segmented from the background prior to image rendering, more informative renderings could be applied, such as shaded surfaces, densitometric projections or some combination of the two (Hany et al., 1998). Segmentation of the intracranial vessels is complicated by the overlap between the vessel intensities and the background intensities. This is particularly true for small vessels.
A variety of MRA segmentation approaches have been proposed, including filter based (Sato et al., 1998), statistical (Wilson and Noble, 1999) and deformable models (Lorigo et al., 2001). Sato et al. (1998) proposed a segmentation scheme based on the application of a line enhancement filter followed by thresholding. The basic idea of this work is that a measure of “lineness” can be computed which would provide better separation between background and vessel voxels, thus facilitating a more accurate segmentation via application of a threshold. In Sato’s particular application “lineness” was estimated using the anisotropies of the Hessian matrix. Because the vessels have a range of widths, a multi-scale implementation is achieved by using different width Gaussians to estimate curvature and then examining the maximum response over all examined scales. Wilson and Noble (1999) developed a statistical model of voxel intensities based on a physical model of blood flow. A modified expectation maximization algorithm was then used to classify time-of-flight (TOF) intracranial MRA images into vessel or brain parenchyma. Lorigo et al. (2001) used a level set approach to segment 3D vascular images. In the level set approach an initial surface evolves iteratively to minimize an energy function which includes image gradient information.
We have developed an alternative vascular segmentation algorithm which provides a sensitive and highly specific segmentation of the vasculature in bright-blood MRA images. This segmentation method, known as the depth buffer or Z-buffer segmentation (ZBS) algorithm (Parker et al., 2000), uses continuity criteria to segment the MIP Z-buffer. This ZBS is a pre-processing step for 3D segmentation, providing an intelligent set of seed points of the 3D vasculature which can be exploited in a variety of manners. To date we have coupled these seed points to a simple 3D region growing algorithm. This segmentation approach makes very few assumptions regarding the nature of the data. The most basic assumption is that contiguous pixels within a vessel region in a MIP Z-buffer image come from more similar depth locations than contiguous pixels in a background region. This weak assumption allows us to segment vascular abnormalities such as aneurysms which violate the line-like geometry assumptions often made in many vascular enhancing filters and segmentation algorithms such as Du et al. (1995), Orkisz et al. (1997), Sato et al. (1998), Frangi et al. (1999), and Frangi et al. (2001).
With the vasculature segmented in a sensitive and specific manner, we are free to perform projections which contain more information than MIP images. For example, densitometric projections and shaded surface renderings can be combined to provide images with both overlap and depth information, as illustrated in Fig. 1. Segmentation will also allow analysis of vessel dimensions and quantitative application of fluid dynamics to determine spatial variations in velocity sheer and other factors that may be indicators of potential disease states. Segmentation may also facilitate computer aided diagnostics.
In this paper we assess the accuracy of the ZBS algorithm in terms of its inclusion of vascular voxels and exclusion of non-vascular voxels. This evaluation is based on a large set of hand classified voxels from intracranial MRA images. In this paper we first review the basis of the ZBS segmentation. Second, we provide a description of our voxel selection method. Third, we examine the ZBS segmentation dependence on a variety of segmentation parameters. Finally, we compare the ZBS algorithm to segmentation based upon vessel enhancement filtering, similar to what Sato et al. (1998) proposed. For this purpose we have used a multi-scale vessel enhancement filter based on an idea proposed by Du et al. (1995).
Section snippets
ZBS segmentation
Here we describe the basic elements of the ZBS algorithm. First, the algorithm segments the MIP Z-buffer to generate an intelligent set of 3D seed points. Second, the 3D seed points are then coupled with a 3D segmentation algorithm. To date this has been a simple region growing algorithm.
Voxel selection
An average of 57.4 (SD=32.2) points were selected at each grid location. Of the 490,256 voxels selected from 14 patients, 449,574 were background voxels, 18,595 were large vessels, 11,244 were medium vessels, 9091 were small vessels and 1752 were aneurysms. Example MIP images showing the classified voxels are shown in Fig. 3. Intensity distributions for the tissue classes are shown in Fig. 4. From Fig. 4 we can see that there is substantial overlap between the classes. However, the means are
Discussion
We have examined the accuracy of the ZBS algorithm for segmenting the intracranial vasculature from high resolution time-of-flight MRA images. This evaluation was based on a large number of voxels (490,256) classified from 14 patients. The algorithm is based on the MIP algorithm. We have demonstrated that whereas the MIP algorithm performance varies with orientation, slab location and slab size in regards to voxel selection properties, the ZBS algorithm performance was relatively insensitive to
Acknowledgements
This work partially supported by NIH Grants RO1 48223 and RO1 53596.
References (19)
- et al.
CURVES: curve evolution for vessel segmentation
Med. Image Anal.
(2001) - et al.
Artifacts in maximum-intensity-projection display of MR angiograms
AJR Am. J. Roentgenol.
(1990) - et al.
Contrast-to-noise ratios in maximum intensity projection images
Magn. Reson. Med.
(1992) - Chapman, B.E., Stapelton, J.O., Christian, B., Hamilton, B., Katzman, G.L., Moore, K.R., Tsuruda, J.S., Parker, D.L.,...
- et al.
Reduction of partial volume artifacts using zero-filled interpolation in 3D MR angiography
J. Magn. Reson. Imaging
(1994) - et al.
Vessel enhancement filtering in three-dimensional MR angiography
J. Magn. Reson. Imaging
(1995) - et al.
Model-based quantitation of 3D magnetic resonance angio-graphic images
IEEE Trans. Med. Imaging
(1999) - et al.
Quantitative analysis of vascular morphology from 3D MR angiograms: in vitro and in vivo results
Magn. Reson. Med.
(2001) Handbook of Medical Imaging
(2000)