- Purpose: There is widespread interest in the segmentation of blood vessels in the brain, for example to study cerebral aneurysms and stenosis. However manual segmentation of veins in an (MR) brain dataset is a cumbersome task and therefore an automated approach is desirable. In 1998 Frangi et al. [1] designed a multiscale vessel-likeliness filter which can automatically segment vessels in CT and contrast enhanced MR angiograms. Recently it was shown that the filter could also be applied to susceptibility weighted images (SWI) in which veins are dark with respect to brain tissue [2]. In this study, an additional filtering technique, vessel enhancing diffusion (VED) [3,4], was applied on a functional SWI dataset [5] in both a low SNR (single functional volume) and a high SNR (averaged volume) scenario. The performance was compared to Frangi-only segmentation with manual segmentation serving as a reference. 2. Methods and Materials: A healthy subject, male 25y, was studied using a Siemens Trio 3T system after informed consent was given. The datasets of the brain were acquired using a high resolution 3D flow-compensated gradient echo FLASH sequence and an 8 channel occipital coil array. The parameters were as follows: matrix size 144x192, 40 slices, resolution 0.75x0.75x0.75 mm3, acquisition time 2 minutes per volume, 8 volumes in total. Since this data was part of a functional SWI study of the human visual cortex [5], only this part of the brain was scanned. The approximate orientation and position of the central slice can be seen in fig. 1a. Due to the nature of the occipital coil array a large signal intensity inhomogeneity is present (fig. 1a-b). Because part of the Frangi filter (see below) is sensitive to (absolute) intensity values, this intensity inhomogeneity was corrected for using a low frequency bias field (fig. 1c). The filters were applied to both a single volume and the averaged volume (after motion correction) to check the performance in a low and high SNR situation, respectively (fig. 1d-e). Manual segmentation was performed on the average volume. 3. Results: Figure 3 shows the maximum intensity projections of the Frangi filter output using three different background noise suppression settings all at the same level and window settings. For both the low SNR dataset in the top as the high SNR dataset in the bottom row, clearly the lowest setting (2%) fails to produce a good contrast between the veins and the background. Even after changing the level and window settings (not shown here) the contrast hardly improved. The 20% setting suppresses the noise well in the high SNR regions, the low SNR regions are still noisy. Increasing the noise suppression further however does not improve the contrast in these anterior regions. 4. Conclusion: VED filtering outperforms Frangi filtering in regions of low SNR, suppressing more noise while leaving the underlying veins intact. In regions of high SNR the smoothing effect of the diffusion approach reduces the amount of detail.
Vein Segmentation from 3D High Resolution MR Venograms by Using Vessel Enhancing Diffusion
P. Koopmans, R. Manniesing, D. Norris, M. Viergever and M. Barth
European Society for Magnetic Resonance in Medicine and Biology 2006.