engleski

Multiresolution CT head image analysis using simulated annealing

Segmentation of computed tomography (CT) head images plays an important role in quantitative analysis of diseases such as human spontaneous intracerebral brain hemorrhage (ICH). The segmentation of the ICH has shown to be a challenging task. Studies involving expert neuroradiologists have shown a variability in manual segmentation of ICH primary and edema regions. A multiresolution probabilistic approach for segmentation of CT head images containing ICH primary and edema regions is presented in this work. In this method, the segmentation problem is viewed as a pixel labeling problem. The pixel labeling problem consists of assigning a label from a pre-defined label set to each pixel in the image. The label set consists of a limited number of possible labels corresponding to segmentation classes. In this particular application the elements of the label set (i.e. the labels) are: background, skull, brain tissue, and ICH. The proposed method is based on the Maximum A-Posteriori (MAP) estimation of the unknown pixel labels (i.e. the segmented image). The MAP method maximizes the a-posterior probability of segmented image given the observed (input) image. Markov random field (MRF) model has been used for the posterior distribution. The MAP estimation of the segmented image has been determined using the SA algorithm. The SA algorithm is used to minimize the energy function associated with MRF posterior distribution function. The problem in the practical realization of the algorithm is that even for a moderately sized image, e.g. $256 \times 256$ the number of variables (i.e. pixels to be labeled) in the optimization space is $65,536$. This number is even larger in case of higher resolution images. This is the reason for the large computational complexity of the algorithm leading to a long execution times. To overcome this difficulty and speed up the algorithm we have developed a multiresolution image analysis approach as follows. In the first step of the algorithm, the SA segmentation is performed on a small-sized image hence reducing the number of variables in the optimization space and decreasing the algorithm complexity.In the subsequent steps at doubled resolutions, results from previous steps are used as initial approximations for the segmentation at next higher-resolution levels. In such a manner, it is possible to reduce the necessary number of SA iterations required to find the final segmentation result. This results in a higher execution speed.

simulated annealing; image segmentation; computed tomography

nije evidentirano