The overall goal of this experiment is to detect midline shifts from brain CT images of patients with traumatic brain injury and thereby estimate intracranial pressure within the brain. In this project, we developed a computational method to analyze the CT images and predict the intracranial pressure or ICP. The incentive for this project is the fact that in many cases where people have head injuries, there is always the suspicion that ICP or intracranial pressure might be elevated, So a ct, we can evaluate swelling of the brain, we can evaluate hematomas, we can evaluate midline shifts or, or tial shifts of the brain, but what we can't cannot predict reliably is what effect that has on the intracranial pressure other than in general, it raises it to a to some degree, and this is where the computational methods that we've been working on are, are helpful.
It can predict with a reasonable level of accuracy exactly what effects the anatomic changes have on intracranial pressure. This schematic diagram presents the overview of the methodology employed in this research. Computerized axial tomography is more commonly known by its abbreviated names, CT scan or CAT scan.
The CT scans are developed from a large series of two dimensional x-ray images taken around a single axi of rotation. In the first step of our approach, the CT scans of the patients with traumatic brain injuries or TBI are acquired and analyzed to detect the midline shift. The next step includes extraction and analysis of other features such as various texture information of CT images and blood amount estimation.
Our approach also uses the extracted features as well as other demographic information to predict ICP machine learning algorithms are used to model the relationship between the ICP and the extracted features. The approximate ideal midline is detected using symmetry of the skull. First using gray scale thresholding, the skull is segmented from the image slice.
Then the algorithm searches every rotation angle exhaustively around the mass center of the skull to find the line that maximizes the symmetry of the resulting halves of the skull. The approximate ideal midline is the line passing through the mass center point with the rotation angle. Maximizing the symmetry, Detecting the ideal midline based on symmetry alone may not always provide us automatically meaningful results.
Therefore, in our method, we take one step further to find specific electronical features such as posterior box theory and the anterior fox attachment to refine the locations of the amine line. Next, the posterior fox cere and the anterior fox attachment to the margins of the sagittal sulcus are identified to refine the approximate ideal midline in order to detect these anatomical features quickly and accurately. Two, searching rectangles are defined based on the approximate ideal midline and its two intersection points.
With the calvarium, the size of the rectangle is chosen to include the anatomical features to be detected. The anterior fox attachment is detected as the peak point of the ridge on the calvarium, whereas the fox cere is detected as the gray line in the posterior region. A bone protrusion is located in the anterior section of the skull.
The fox cery extends from this point. This anatomical feature can be used as a starting point for the midline. It can be seen that the protrusion curves down to a local minimum point at the anterior edge of the foxer.
If one views the lower edge of the skull bones as a curve on the image plane, then the curve can be modeled as a one dimensional function. Detection of the protrusion point becomes a matter of finding the minimum of a sampled one dimensional function. With the input of the upper segmented skull bone, the first step is to extract the lower edge of the bone.
The next step involves the detection of the protrusion using the local mini. Here, the local minimum also represents the endpoint of the protrusion. This schematic diagram shows how the fox at the lower part of the brain is detected.
Since the aim is to detect light gray lines and not the darker lines in the first step, the median value of the area is used as a threshold. Using this threshold, all the gray scale values below it are set as the median value. Then a zero crossing edge detector is used to generate the edge map.
This maintains richer edges during sobell detection and allows removal of unwanted edges in the next step. The edge map is then refined step-by-step using known features of the fox cere. This is first done by refining the edge map using intensity and gradient following which the size threshold and concentration threshold is used to refine it.
Further here, the concentration threshold is nothing but the density of the edge points. Finally, Hof transform is applied to detect lines representing the fox.Sarah. The results of the HOF transform are usually a set of lines.Two.
Constraints must be established to extract the desirable lines from this set. First, the angle of the line must be in the range where the lines are concentrated. This range is obtained by calculating the statistics of the angles of the detected line.
Second, the line must lie inside the range where the cluster of lines are concentrated. The final ideal midline at this stage is chosen as the longest lines satisfying the above constraints. This image shows the two regions of interest and refined ideal midline based on the inner protrusions.
The green line is the refined ideal midline, and the red line is the approximate Position of the ideal midline of the rain before refinement. This image shows the 3D model of the ventricular System with the presence of an MRI slice of the brain. The red portion of the image represents the ventricular system in 3D.
The ventricular system consists of two bilateral ventricles on the top, the third ventricle in the middle, and the fourth ventricle at the bottom. The reason we extract information about a ventricle is because when the pressure is changed inside the brain due to an injury, it deforms and therefore is a suitable candidate to measure the deformation of brain tissue and pathological cases. In successive MRI or CT image slices, the ventricular system appears in various shapes on the right, the extracted shapes of the ventricular system From an MRI are shown.
The aim of the CT segmentation is to identify these shapes from the original CT scans. In our proposed method, the segmentation process is separated into two parts. First, an initial low level segmentation method is applied to group pixels into different parts.
Then a high level template matching method is used to identify ventricles from segmented results. This image illustrates the low level segmentation based on a Gaussian mixture method. First, the original CT images filtered using median filtering.
The result is shown in the left image. Afterwards, the K-means algorithm is applied to obtain a core segmentation. This result is shown in the middle.
After the K-means algorithm is applied, goss and mixture method is then applied. The image on the right shows a different goss and distributions on the 2D image, which are initialized. Using the came segmentation expectation maximization is used to evolve the parameters and finally, attuned segmentation result is obtained.
The ventricular system in the Z direction does not show significant variation across the training set. Therefore, mappings can be approximated with linear form. In the method described here, these mappings are first initialized manually and then optimized using a set of training images.
There are three constraints to accept the initial ventricle segmentation. Number one, the segment needs to be relatively large. Number two, the segment is not close to the edge of the brain.
Number three, the segment intersects with the ventricle template. The figures at the bottom show an example of the ventricle recognition step from the results. One can see that the ventricle parts are successfully recognized Using size, bounding box, and template constraints.
In order to estimate the actual midline, we first match the segmented Ventricles against the templates by shape matching. This figure shows the process of the matching. First, the edge points are sampled.
Then by optimizing a matching cost of the two shapes. The correspondences of the points between the two shapes are made. For example, in the second row, the last image shows the correspondence by connecting the red points from one shape to the green points in the other shape.
Because the ventricular system is a 3D structure and has different shapes on different CT slices, we perform the shape matching across all CT slices. For each slice, we define the feature points that can be used to estimate the midline. For example, the feature points in the lateral ventricle shapes are the inner edge points of the two lateral ventricles.
These feature points are identified by their correspondence to the feature points on the template, which are labeled manually before the matching process. Here we show many matching results by choosing different templates. For each shape matching, there is a match cost.
By selecting the minimal match cost, we find the best template to be used in the shape matching and identify feature points. In this example, the points from the segmentation results are shown in green, and the points from the templates are shown in red. By comparing the match cost, we can see that the first template is the best to use.
Once the feature points are identified, we calculate the X coordinates of these feature points. Usually, there are feature points on the left side and feature points on the right side. The X coordinate of the actual midline is estimated as the average of the X coordinate of the left and the right feature points.
Here we show four estimation results. The green vertical line represents the estimated X coordinate of the actual midline. The match template is shown in the blue points.
The identified feature points are shown in blue. From the results, we can see that the estimation of the actual midline works very well after the actual midline is estimated. The calculation of the horizontal midline shift S is straightforward.
X ideal is the X coordinate of the ideal midline, and X actual is the X coordinate of the estimated actual midline. The images shown here Illustrate both the estimated ideal midline and the actual midline. Besides using midline shift, these features that we extract may provide us additional information in predicting ICP levels.
The main idea behind this is to extract as many features as possible, and then by using feature selection methods, we are able to retain only those features that are relevant in providing ICB information. The features extracted from CT scans include midline shift blood amount, and texture patterns. Features from other sources include demographic information such as patient age, trauma score, and injury severity score.
A gian mixture model based segmentation algorithm is used to label the pixels. Pixels are grouped into four categories, blood cerebral fluid, gray matter, and white matter texture. Patterns in the CT images may include indications of the state of the brain.
High levels of intercranial pressure may change the texture pattern. We focus mainly on the texture patterns in the area without cerebral spinal fluid or blood texture analysis is applied on small sub images or windows of the CT image. Six windows are selected in each CT image.
Texture features are extracted using discrete foer transform and discrete wavelet transform. The long and short of it is there is so much information available. It becomes very problematic for physicians no matter how experienced they are to process all of these signals at the same time.
So the current project, which aim to look at the CAT scan and do automated imaging would be great use to emergency physicians and traumatologists like myself to help us to process this data much more quickly. The dataset has 17 patients with mild to severe traumatic brain injuries. Each patient has several CT scans, and there are 57 total scans used in the study.
A value representing intercranial pressure is recorded every hour. There are two categories of ICP level elevated. ICP is defined as ICP greater than 12 tor.
For normal ICP, it is defined as ICP less than or equal to 12 tor. In this dataset, there are 33 normal cases and 24 cases of elevated ICP. In order to create the model, we use two stages of tenfold cross validation.
The first is a nested stage in which feature selection is performed. In the second stage, a genetic search method is used and the classifier used in conjunction With the search method is support vector machine. This image shows the ideal Midline estimation.
The red line is the approximate ideal midline based on the symmetry of the skull. After refinement using anatomical features, we have a better estimated ideal midline, which is the green line. In the image, these images show the estimated actual midline.
The green lines are the estimated actual midline, and the blue points are matched templates. The image on the right shows the estimated ideal midline to the left of the center and the actual midline. On the right, we can see the resulting midline shift from the estimation.
Here we also present the quantitative results in the evaluation of our methods. In most of the CT slices in our dataset, the error between the ideal midline estimated by our methodology and the physician labeled midlines is around two pixels or one millimeter. For the actual midline, more than 80%have less than 2.25 millimeter difference under certain quality control of the segmentation results.
The intercranial pressure prediction evaluation has about 70%Accuracy using tenfold cross validation. First of all, we separate the low level and high level Segmentation processes. Although this design can benefit from incorporating different algorithms, we may not exploit all information in the low level segmentation.
By combining the low level segmentation and high level segmentation, we may have better segmentation results. Following this idea, we can use methods such as model based segmentation or registration based segmentation in order to have better prediction, as well as providing more evaluation of the proposed method. Collecting more data sets To have a larger sample size will benefit the study.
