Master Thesis Segmentation Tool Features

		Back to Main Back to Master Thesis Features :: Results :: Conclusion :: Downloads

The features of the developed tool are:

• A DICOM Parser reading the data which stems from an Electron-Beam CT. The dynamic data sets comprise 10x8 ECG-triggered images (10 images in time at each of 8 slice locations). Each image consists of 360x360 pixel which stand for a real distance of about 0.8 mm, respectively. Resolution in z-direction is rather limited with its 8mm slice thickness. Images are taken in long-axis configuration.

CT image acquisition setup	Heart long-axis configuration

• Left ventricle volume estimation by using a parametric model based on an ellipsoid. It is the so-called Three-Axes Method by Greene. Therefore the projection with the largest area of the left ventricle has to be located from the data set. There a long and a short axis of an ellipse are measured and the volume is calculated by a modified formula for the ellipsoid volume.

Illustration of the Three-Axes Method by Greene

• The following block diagram gives an overview of the volume estimation techniques based on digital image processing which were investigated in this work

Image Processing Block Diagram

• Image Preprocessing like automatic cropping of the images, scaling to a meaningful gray-level range and performing a median filter for noise suppression.


• To find the left ventricle area in the images an automatic identification of a region of interest was implemented which took temporal information into account. Subtracting temporally adjacent images and adding thresholded results led to promising ROIs.


• For fully automatic detection of the left ventricle projection with the largest area a Hough Transform for ellipse detection was implemented. This algorithm is based on geometrical ellipse properties and tries to detect the largest ventricle area found in a data set provided the information from the automatic Region of Interest algorithm.


• For the actual segmentation in order to evaluate left ventricle volumes the following three approaches were compared:
  
  
  • A thresholding algorithm based on the histograms of the images. This algorithm performed very poor due to its inherent lack of handling complex tasks.
  
  
  • An active contour algorithm (Snakes) was implemented. This algorithm tries to minimize an energy term of a polygon by looking at internal energies like curvature and bending of the polygon and on the other hand image gradient information. By balancing these two terms the snake is attracted to certain regions. This algorithm only showed average results, due to the snakes problem of proper initialization and parameter-tuning.
  
  
  • The semi-automatic LiveWire segmentation paradigm was implemented. This algorithm is based on graph theory and dynamic programming. A minimum-cost path map according to image gradient information is calculated and the user has the option to move around with a mouse cursor to roughly delineate the segmentation of an object. The tedious accurate segmentation is performed by the algorithm.