before the experimental results from a 79 year all patient is presented. Finally, some concluding remarks are stated.

The human brain weighting about 1300-1400 grams is immersed in cerebrospinal fluid, denoted CSF.

This fluid is produced by special cells in parts of the ventricular system, inside the brain at a rate of about 500 ml/24 hr. It is drained into the venous system of the brain and thus circulates within the ventricular system, over the brain hemispheres and around the spinal cord in the subarachnoid space. Normally, there is a steady state of production and draining such that the ventricular system contains about 10-50 ml and the subarachnoid space 120-220 ml, dependent on age and brain size, among other biological variation.

CSF is of vital importance for maintaining healthy environment for the brain cells. It also reduces stress and strain on the brain during head movement since the weight of the brain in situ becoms only 45 grams.

Changes in CSF volume and its spatial distribution are important signs associated with several neurological diseases. Of particular interest is the condition of hydrocephalus with increase in CSF volume within the cranial cavity.

This condition can be present at birth and is then called congenital hydrocephalus. The neurosurgical procedure of by-passing and shunting CSF directly into the venous system can then be necessary.

The task in our case is to segment CSF of the ventricular system in hydrocephalic brain using multispectral 3D MRI data, and perform quantitation and rendering of this fluid space.

If this can be properly achieved it will be an important aid in diagnosis and follow-up of the aforementioned diseases. Of particular interest is quantitative evaluation of shunting operations in children and in the eldery. Also, brain volumetry is important to many research problems in cognitive neuroscience.

In this study we have used a Siemens Impact MR scanner operating at 0.95 Tesla. We used the standard circularly polarized head coil and the scanner was equipped with 15.0 mT/m gradient strength allowing fast 3D pulse sequences.

Each 3D channel image consisted of:

128 contiguous slices from a 180 mm slab covering the head using 128 x 256 acquisition matrix with sagittal slicing from left ear to right ear.

With a 256 x 256 display matrix the in-plane resolution was about 1 square mm, and the voxel volume was about 1.4 microliter.

Three different gradient echo 3D pulse sequences were used:

Time of acquisition (TA) for each 3D channel was between 6 and 7 minutes.

This slide illustrates the tissue classification method which was used.

It can be divided into a training part and a classification part.

Training, requiring manual intervention, was restricted to one central sagittal slice, while the remaining 127 slices were classified from this design set.

In the training procedure we used a combination of unsupervised k-means clustering and manual labelling of the resulting clusters with k=9 (Taxt et al.,1992; Taxt and Lundervold, 1994).

As data models we used the multivariate Gaussian distribution to describe each tissue. This was based on previous experience and inspection of the training sample distributions of signal intensity in each channel for each tissue category.

We used a contextual classifier with first-order 3D neighbourhood in a Bayesian framework with 0-1 loss and uninformative prior probabilities.

In the estimation of tissue-specific multinormal probability densities, we used the sample mean vector and sample covariance matrix.

In this application we distinguished between 7 different tissue categories.

As contextual classifier we used a 3D generalization of the Haslett Markov random field model
( Haslett, 1985)

The Markov property is here expressed in the upper equation.

The contextual terms are introduced as products over the first-order neighborhood in 3D ( Left, Right, Front, Back, Up, Down ) incorporating summed products of transition probabilities between neighboring voxels and tissue-conditional densities.

No estimation of transition probabilities was done, only plain specification.

To obtain 3D segmentation of CSF we applied a recursive 3D connected component labelling algorithm, published by Cline and coworkers (Cline et al., 1987).

Again, we used the first-order neighborhood in the 3D lattice.

Volumes were calculated by counting voxels in connected CSF components within the lateral ventricles.

Finally, 3D renderings were performed by combining a subset of the 3D FLASH data and the complementary volume consisting of the segmented lateral ventricles using the Stanford VolPack volume renderer (Lacroute and Levoy, 1994 - http://www-graphics.stanford.edu/).

This slide shows the slice used for training.

Even if we wanted to distinguish between 7 tissue categories we used 9 clusters in the k-means clustering. Three clusters represented fluid due to the heterogeneous signal from flowing CSF in the lateral ventricles, as seen in the T2-weighted channels DESS and FISP.

Using VolPack software, we computed renderings from different viewpoints.

We can clearly see the enormously enlarged lateral ventricles, where voxelcounting in the 3D classification gave a volume of about 550 ml - which is an order of magnitude larger than normal values for this structure.

The same method has also been applied to normal individuals to segment lateral ventricles and subarachnoid space ,

as well as brain parenchyma.

In conclusion:

Quantification and visualization of cerebrospinal fluid spaces in the brain is clinically important and technically possible using multichannel 3D MR imaging and multispectral tissue classification.

However, extensive clinical validation is needed before this approach can come into routine clinical use.

In this respect, more optimal pulse sequences should be investigated (such as splitting the two echo signals in the DESS sequence).

Furthermore, better use of prior information and more valid data models should be explored.

Also, use of partial volume classification could improve accuracy of the volume estimates as well as the renderings of the fluid spaces and surrounding tissue.

<- MIAPR home