Jonathan Arvidsson^{1,2}, Oscar Jalnefjord^{1,2}, Fredrik Kahl^{3}, Magnus Båth^{1,2}, and Göran Starck^{1,2}

Inclusion of voxels containing CSF and/or blood vessels can bias ROI statistics used in DSC-MRI analysis. In order to address this problem we propose an automatic method for tissue segmentation based on hemodynamic features obtained from DSC-MRI data. Application of the method in test subjects shows promising results.

Dynamic susceptibility contrast (DSC) MRI can be used to characterize the microvasculature of the brain. A commonly used analysis approach is to extract a summary statistic from a region of interest (ROI) in the brain. However, if voxels containing other tissue types such as cerebrospinal fluid (CSF) or blood vessels are included in the ROI, the statistic will be biased. A segmentation that identifies voxels of unwanted tissue types would thus be beneficial. Furthermore, the segmentation should ideally be based on images already included in the examination to reduce examination time.

A segmentation method using the DSC data itself for identification of CSF
voxels has been proposed by Kao et al^{1}. In this study, that
method is extended to also identify vessels by addition of another feature
extracted from the DSC data in combination with regularized clustering.

The method
for identification of CSF voxels proposed by Kao et al. exploits the long T1 of
CSF and the fact that steady state has not been reached during the first
dynamics^{1}. The feature, called “first ratio”
(FR), is the signal at the first dynamic normalized to some reference. In this
study, FR was calculated as

$$FR=\frac{S(1)}{S_0}$$

where S(t) is the
signal at the t^{th} dynamic and S_{0} is the average signal at
steady state before contrast agent injection, i.e. during baseline.

Vessels can be assumed to differentiate from other tissue types by having a higher increase in contrast agent concentration after injection. To capture this characteristic, a feature based on the peak concentration (PC) is proposed, calculated as

$$PC=\frac{1}{3} \sum^{T_p+1}_{t=T_p-1}C(t)$$

where C(t) is the
contrast agent concentration at the t^{th} dynamic and Tp is
the dynamic with maximum C(t), i.e. the peak concentration. Conversion from
signal to concentration was performed according to

$$C(t)=-\frac{1}{TE}\ln\frac{S(t)}{S_0}$$

where TE is the echo time.

The k-means
algorithm^{2} was used to cluster voxels into
three clusters corresponding to CSF, vessels and brain parenchyma. To improve
interpretability and robustness of the clustering, a regularization term was
added to the k-means objective function

$$J(\mu_k)=\sum_k\left(\sum_{x\in A_k}||x-\mu_k||^2+\lambda_k||\mu_k-\mu_{0,k}||^2\right)$$

where x is
the feature vector of a voxel, A_{k} is the set of voxels in the k^{th}
cluster,
μ_{k}
is the center of the k^{th} cluster and μ_{0,k} is the center of the k^{th} cluster
proposed prior to the optimization. λ_{k} are regularization factors calculated as

$$\lambda_k=\frac{\sum_k\sum_{x\in A_{0,k}}1}{\sum_k\sum_{x\in A_k}1} $$

where A_{0,k} is the set of voxels
in the k^{th} cluster proposed
prior to the optimization.

A_{0,k}
and μ_{0,k} were estimated separately for the CSF and
vessel classes. For the
CSF prior class, Otsu thresholding^{3} of the FR feature was used^{1}. Furthermore, connected components
(4-connectivity) containing fewer than 10 voxels were excluded from the class. For
the vessel prior class the 10 % voxels with highest PC value was chosen. Voxels
assigned to both the CSF and the vessel prior classes were excluded. Voxels not
included in neither the CSF nor the vessel class were assigned to the brain parenchyma
class followed by the same connected component reduction step as for the CSF
class. Some voxels were thus not necessarily included in any of the prior sets
A_{0,k}. All processing was done using MATLAB.

To study
the usefulness of the proposed method, MR images of two idiopathic normal
pressure hydrocephalus patients were acquired with a 1.5T Philips Gyroscan
Intera. For anatomical reference FLAIR images were acquired (voxel size 1.2×1.2×3mm^{3}).
DSC-MR images were obtained every 1000ms with GE-EPI (voxel size 1.8×1.8×5mm^{3}).
A bolus (5ml/s) of 0.1mmol/kg body weight Gd-DTPA (Schering, Magnevist 0.5mmol/ml)
was initiated at the tenth dynamic into the right antecubital vein. Skullstripping
was employed to form an ROI encompassing the brain. For each subject, all voxels within
the ROI were extracted and clustered using the described method. Before clustering, features were
normalized to have zero mean and unit variance.

- Kao YH, Teng MMH, Zheng WY, Chang FC, Chen YF. Removal of CSF pixels on brain MR perfusion images using first several images and Otsu’s thresholding technique. Magn Reson Med 2010;64:743–748.
- Macqueen J. Some methods for classification and analysis of multivariate observations. Proc Fifth Berkeley Symp Math Stat Probab 1967;1:281–297.
- Otsu N. A Threshold Selection Method from Gray-Level Histograms. IEEE Trans Syst Man Cybern 1979;9:62–66.