Segmenting Linear Parts using Layered Region Growing

The segmentation of linear structures in a gray-scale image is widely used in a variety of image analysis and visualization tasks. In case of biomedical image analysis of volumetric data (such as CT or MRI scans), it is very common that a linear structure (blood vessel) is attached to a sheet-shaped (planar) and/or blob-shaped (ellipsoid) region (large tissue or fluid mass such as cistern basin in the brain) but where the image intensity of both structures is the same. Segmentation algorithms (threshold based or statistical approaches) are unable to clearly define and therefore segment the areas of such attachment. Automatic detection of the attachment and locating the start and end of linear structures is important for visualization and quantification in biomedical analysis. In this paper, we present the layered region growing (LRG) algorithm to segment the linear structures from volumetric data. First, we define the LRG algorithm for volumetric data, then present a layer frame for extracting the region of interest, and introduce a width function on the layer. Next, we show that the width function is of constant order if LRG is used to extract a linear structure whatever the selection of seed, O(n) for a sheet-shape, and O(n) for a blob-shape, where n is the number of layers in the region. Last, we give an automatic method to detect the change of the width function as a criterion to stop the region growth at the attachment areas where different shaped structures meet. Promising results are shown on simulated data as well as on cerebral vascular CT data.