Segmentation of spinal cord images by means of the watershed and region merging together with inhomogeneity correction

Abstract

The paper presents a morphological method for segmentation of high field Magnetic Resonance (MR) images of the human spinal cord and extraction of the gray matter mask. These images are of low quality and poor contrast. The inhomogeneity of brightness in the image is usually more pronounced than the difference in brightness between the gray matter and the white matter. Due to this inhomogeneity, it is very hard to use watershed segmentation for automatic extraction of the gray matter, and what remains is manual pointing out of a hundred or more regions belonging to the gray matter. However, as shown in the paper, by using the White Top Hat (WTH) transform with a large structuring element, one can correct the images, significantly reducing the inhomogeneity and appropriately modifying individual region statistics. In particular, watershed segmentation is carried out on the original image, whereas region statistics used for region merging are calculated from the corrected image. Then the extraction of the gray matter mask is carried out in a semi-automatic way, with the user pointing out the first region belonging to the gray matter area, and the program selecting subsequent neighboring regions based on the statistics of the regions. The method was tested on images coming from different cross-sections of the spinal cord, and the results indicate that the process of extracting the gray matter mask has been significantly speeded up and improved.