25-05-2012, 12:26 PM
analysis of medical images
analysis of medical images.doc (Size: 1.15 MB / Downloads: 33)
INTRODUCTION
Often the analysis of medical images for the purpose of computer-aided diagnosis and therapy planning contains segmentation as a preliminary stage for the visualization or quantification. For medical CT images, many methods were recently employed for segmentation, for instance interactive thresholding aided by morphological information, region growing and region splitting and merging, active contours, the use of cluster analysis methods, or watershed transformation. In the individual segmentation methods different complex models of the apriori information about the expected contents of the image are used. The applied apriori knowledge consists of a combination of anatomical/physiological information and of information about the image formation process. The more the contributed model information is verifiable, the more likely is the possibility of automation of the segmentation process.
Also, the result of the algorithm can be predicted easier. Especially in medical segmentation tasks the model information is often too complex or not exact to specify so that a completely automatic extraction is not possible. The aim of the development of a segmentation method should be to minimize the part of the interactively introduced model information and to maximize the part of the automatically analyzed model information. Model information should be supplied interactively only in those cases where it represents knowledge that is readily available to the user and where it can be entered in a robust fashion. Contravening this rule may cause reactions of the process to user interaction which are perceived as being inconsistent with the user’s expectations.
OVERVIEW OF THE ALGORITHM
The algorithm was implemented in Matlab. The automated algorithm begins with the analysis of the gray-level histogram for a 3D volume and continues by iterating through each section. Analysis of the histogram, together with thresholding and morphologic operations, results in an initial estimate of a liver boundary. Information from adjacent sections is used to modify the estimated boundary. The boundary is then further refined by using edge information in the image. This final step is achieved by first representing the boundary with use of the parametrically deformable contour model, and then matching the model to an edge-enhanced image through optimization. User modification of the output boundary can be made after a single section has been segmented, or it can be applied after all sections have been segmented by the automatic algorithm. By using both gray-level distribution analysis and edge detection, it is reasonable to expect the algorithm to perform better than use of either method alone.
DERIVATION OF THE HOMOGENEITY MODEL
The objective of our work is the development of a simple and robust possibility to describe the homogeneity criterion in a simple form. This description method is to be used for segmentation of 2D and 3D structures in CT data. Our knowledge of the homogeneity model is based on knowledge about the image formation process. In CT images values of the image represent average x-ray absorption distorted by noise and artefacts. The absorption itself is assumed to be constant for a given anatomical structure. The noise is assumed to be zero mean Gaussian noise with an unknown standard deviation. For CT images, the main artefact is assumed to be the partial volume effect (PVE). Beside noise artefacts and partial volume effect, other problems like shading effects complicated segmentation.
Homogeneity without visible shading may be defined as likelihood of belonging to a Gaussian distribution of gray values with a given mean and standard deviation. The mean is the absorption value of the tissue in CT images. The standard deviation accounts for variations due to noise. The effects of the PVE can be captured in an approximate fashion by assuming different “standard deviations” for gray values that are higher or lower than the mean.
GRAY-LEVEL THRESHOLDING FROM ANALYSIS OF HISTOGRAM
The first step in our automated technique is analysis of the gray-level histogram for the whole volume and thresholding of each section. Representative histograms of a CTAP scan and a contrast enhanced CT data set are used. The peak corresponding to the liver in each of the images is indicated. This peak position can be located by using the technique developed by Heath et al, which automatically searches for the liver peak by exploiting information that concerns both the amplitudes and the gray-level range of the peaks. The two thresholds for the liver were determined by two offsets from the peak based on the standard deviation (SD) of the gray level of the liver. On the basis of our observation and that of Woo, the intra case gray-level SD of the liver is approximated. The initial lower and upper thresholds for CTAP images are chosen as less than and greater than the peak value, respectively. For contrast-enhanced CT images, the lower threshold is 3 SDs less than the peak value and the upper threshold is 2 SDs greater than the peak value. Kidneys are usually less attenuating (have lower CT numbers and thus appear darker) than the liver on CTAP images and the kidneys are usually more attenuating (brighter) than the liver on contrast-enhanced CT images.