Author(s)

J. L. Foo & E. Winer

Abstract

This paper presents an approach for segmentation of digital medical images using a multi-phase probabilistic approach. The first phase of the approach is enhancement of the image data using an unsharp mask sharpening algorithm. This vastly improved the clarity of the images prior to segmentation. The segmentation of an object within the image is then achieved through two more phases. The first phase is a thresholding process, where each pixel is scored based on a similarity criterion for a chosen seed point. Pixels with a score satisfying a required minimum are selected into the segmented region. A second phase is then instituted for pixels not initially selected. This phase involves a probability selection, based on a Monte Carlo simulation. A probability for each pixel is formulated based on the pixel’s intensity and location. This probability is them compared against a random probability to determine if the pixel is included in the segmented region. To facilitate the processing of multiple image slices, an automated relocation algorithm was developed to move determined seed points from image to image as the shape, size, or location of the object (i.e. organ) changes. The concepts and techniques developed were tested on three separate medical studies with one being shown in this paper. The results showed that the image enhancement process outlined features and details within the data that were not previously apparent. The segmentation process extracted the desired object more completely when compared to other segmentation techniques. Keywords: image segmentation, probabilistic, region growing, thresholding, Monte Carlo, medical imaging, DICOM.

Keywords

image segmentation, probabilistic, region growing, thresholding, Monte Carlo, medical imaging, DICOM.