Processing Image Segmentation

Automatic segmentation of CT images admittedly presents significant challenges in computer vision [42]. The primary reason is that the organs are flexible and their size and shape varies as a function of patient characteristics and imaging parameters. Organs are usually accurately localized on CT slices (Fig. 4.4.) but the detection and separation of their boundaries from those of their neighbors and the background is often a difficult task due to the obscure, fuzzy, and irregular edges that are often superimposed by other structures [33, 42]. Even human experts have difficulty in providing unambiguous outlines of the organs' boundaries and consequently present significant inter- and often intraobserver variability, the magnitude of which is a function of experience and training. Historically, standard techniques, such as absolute thresholds, edge detection, and region growing algorithms that perform some type of operation on the gray level distribution of the image pixels are not, by themselves, sufficient for CT segmentation. Combinations of modules, as the one shown in Fig. 4.5, and advanced approaches, e.g., knowledge-based segmentation [42], are necessary to solve this problem.

Several methodologies are reported in the literature for CT slice segmentation although not necessarily focused on pancreatic tissue or pancreatic tumor segmentation. Methods proposed for organ segmentation in CT slices include pixel based (thresholding), edge based, region based, and clustering methods

[43]. Interactive segmentation of various organs has also been proposed for 3-D visualization. The reported work used simple thresholding and morphological operations that were interactively controlled by a human user via a 3-D display

There are several free software packages that can be used for the segmentation and registration of CT slices. One of them funded by the National Library of Medicine (NLM) is the Insight Segmentation and Registration Toolkit (ITK) and can be downloaded from ITK is open-source software that was developed jointly by six principal organizations to support the Visible Human project of NLM. ITK includes several basic segmentation and registration techniques that have been implemented for a variety of medical image analysis applications. In this work, we experimented with several of the methods implemented in ITK. Particularly, region based, threshold select, geodesic active contour segmentation, and fuzzy connectedness with Voronoi classification

Itk Medical Image Segmentation Purpose
Figure 4.6: (a) Original helical, contrast enhanced CT slice with a tumor at the head of the pancreas indicated by black arrow. (b) Region based segmentation using ITK software on Fig. 4.6(a).

were some of the techniques tested for the segmentation of the pancreas and pancreatic tumors. Initial results suggested that region growing was the best approach because most of the other techniques clustered the majority of the structures in the image together not allowing separation of the pancreas from the other organs. But even with region growing, the pancreas and associated tumor could not be separated from the liver if the pancreatic structures were to remain in the segmented image; separation occurred at the expense of losing most of the information from the gland and associated tumor. Representative segmentation outputs from the region growing approach of ITK are shown in Figs. 4.6(b) and 4.7(b) for two CT slices that contain a mass at the head (Fig. 4.6(a)) and tail (Fig. 4.7(a)) of the pancreas respectively. It should be noted that, although not fully optimized for this application, the tools included in ITK are not likely to yield, by themselves, the desired segmentation outcome because of the low contrast differences between adjacent organs and the way region growing operates. The initial problems we identified in the application of conventional segmentation techniques on CT images of the pancreas include the following:

1. Gray tone segmentation algorithms do not produce accurate regions of the target organ. This is because two different regions of the pancreas or two different organs can have the same or similar gray level tones in CT

Itk Segmentation
Figure 4.7: (a) Original helical, contrast enhanced CT slice with a pancreatic tumor at the tail of the pancreas indicated by white arrow. (b) Region based segmentation using ITK software on Fig. 4.7(a).

images. Hence, differentiation based on gray level alone is not likely to yield consistent and robust results.

2. The shape of the various organs in the CT slices is not always well defined or consistent from slice to slice. So, it is difficult to select generally applicable characteristics. CAD development is likely to require an adaptive process to deal with this variability.

3. Thresholding techniques based on single global values are not likely to succeed because the gray values of the organs are case-dependent. Gray values depend on the chemical contents of each organ and the physical condition and characteristics of the patient. Gray level normalization may provide a solution to this problem but should be done consistently across slices within the same scan so that it does not prevent registration and reconstruction processes. It should also be done with consideration of the variations among different cases, pathologies, and image sources.

Despite limited performances, however, some of the conventional segmentation techniques, including those implemented in ITK, could be used in the first segmentation step for external signal removal and/or removal of uninteresting structure(s) within the slice, e.g., spleen or kidneys, and isolation of major organs including the pancreatic areas. This could make the job of subsequent segmentation steps easier and more successful.

0 0

Post a comment