hierarchical segmentation of ct head images - Semantic Scholar

1 downloads 0 Views 104KB Size Report
4 J. G. Thomas, R. A. Petters, and P. Jeanty, Automatic seg- mentation of ultrasound images using morphological operators",. IEEE Transactions on Medical ...
1

HIERARCHICAL SEGMENTATION OF CT HEAD IMAGES Sven Loncaric and Dubravko Cosic and Atam P. Dhawan

Faculty of Electrical and Computer Engineering, University of Zagreb Unska 3, 10000 Zagreb, Croatia, E-mail: [email protected], [email protected] 

Department of Electrical and Computer Engineering, University of Cincinnati Cincinnati, OH 45221, USA, E-mail: [email protected]

Abstract | Quantitative analysis of head images obtained by computed tomography (CT) requires accurate segmentation. A new method for automatic segmentation of human spontaneous intracerebral brain hemorrhage (ICH) from digitized CT lms is presented in the paper. The proposed segmentation method has a two-level hierarchical structure. The segmentation at both levels is based on the unsupervised fuzzy C-means (UFCM) clustering algorithm but using di erent feature vectors. At the higher hierarchical level, clusters obtained by UFCM algorithm consist of several disconnected image regions. An image labeling algorithm labels each image region with one of the following labels: background, skull, brain, and ICH to obtain the global image segmentation. At the lower hierarchical level the brain region is further segmented using UFCM clustering to obtain the edema region and the ventricle region. The advantage of the method is that it is not sensitive to variability in image brightness. The method has been tested on real CT images and has performed correct segmentation. I. Introduction

Intelligent biomedical image analysis has become a powerful tool for study of internal structures of the human body. Methods for automatic segmentation enable accurate 2-D and 3-D quantitative analysis and measurements with minimal intervention from the human expert. Quantitative analysis of human head images obtained by computed tomography (CT) requires accurate segmentation of CT images. Several approaches for segmentation of brain images have been proposed in the literature such as methods based on neural networks [1], [2], morphological processing [3], [4], expert systems [5], [6], and clustering methods [7], [8]. A new method for automatic segmentation of CT head images of human spontaneous intracerebral brain hemorrhage (ICH) based on clustering and labeling techniques is presented in this paper. II. Methods

The proposed segmentation method has a two-level hierarchical structure and is based on unsupervised clustering and labeling techniques. The method is hierarchical in the sense that the segmentation is performed rst on a global level dividing the image into a number of major regions such as background, skull, brain, and ICH. In the subsequent step, at the lower (local) hierarchical level, the brain region is further divided into local subregions such

as edema and ventricle regions. The two hierarchical segmentation levels are described in the following sections. A. Global segmentation

At the global level a combination of clustering and labeling algorithms is used to perform segmentation of the CT image. The unsupervised fuzzy C-means (UFCM) clustering algorithm [9] with pixel brightness as a feature has been used to generate three clusters having uniform brightness. The three clusters correspond to dark, medium, and bright areas of the image. Each cluster consists of several disconnected image regions. Although a number of clusters is known, the unsupervised algorithm is used because it avoids the initial guess problem present in supervised clustering algorithms. The UFCM algorithm has been modi ed to use an alternative mechanism for new cluster center selection. After the image is partitioned into the three clusters an image labeling algorithm assigns a label from the prede ned label set to each of the image regions. The labels are: background, skull, brain, and ICH. Multiple constraints including region-neighborhood and region-label relations are imposed on the labeling algorithm to narrow the solution space. Discrete relaxation and the backtracking tree search algorithm with forward checking [10] has been used to nd the solution of the region labeling problem. B. Local segmentation

At the lower level of the hierarchical algorithm the brain region is further segmented to obtain the ventricle region, the edema region, and the rest of the brain. The ventricle region is much darker than the other parts of the brain and for that reason is easy to segment. On the other hand, there is a lot of ambiguity in the edema region localization, as opposed to the primary ICH region which is well localized. The edema region may even not be present in some cases. The ventricle region is segmented by means of the UFCM algorithm using the pixel brightness for onedimensional feature vector and two clusters. The two clusters correspond to the ventricle and the non-ventricle regions. Segmentation of the edema region is obtained by means

2

Fig. 2. Local segmentation results: small, medium, and large edema regions.

heuristics and parameters usually present in segmentation algorithms. The method is not sensitive to variability in image brightness due to CT lm development process and Fig. 1. Global segmentation results: original image, background, di erent CT scanners. skull, brain, and hemorrhage (white regions).

of the UFCM algorithm and using a two-dimensional feature vector. The rst feature is the Euclidean distance between a pixel and the ICH. The second feature is the ratio between a pixel value and the average pixel value in a neighborhood of the pixel. The cluster which represents the edema region is easily recognized because the edema region is the closest region to the ICH and is darker than the neighborhood. The edema segmentation has shown to be a challenging task even for human experts because it is not well localized. For that reason, additional exibility is provided by allowing a user interaction for selection of the optimal segmentation result among several variable-sized segmentation candidates. The candidates are obtained using variable neighborhood sizes of 5  5, 7  7, and 9  9 in the computation of the second feature. III. Results and Discussion

The algorithms have been implemented in C language. The proposed method has been tested using real brain image data with a size of 256  256. The CT images were obtained by digitizing the CT lms using a CCD camera. The original CT brain image and the four global segmented regions are shown in Figure 1. At the lower hierarchical segmentation level the brain region is further divided into edema, ventricle, and the rest of the brain. A variability in the size of the segmented edema regions is achieved using a variable-sized masks in computation of the second clustering feature. The three obtained edema regions are shown in Figure 2. The segmentation time on a SUN Sparcstation 2 is several minutes. Automatic segmentation of CT head images is a dicult problem. The proposed technique performs segmentation of background, skull, and ICH regions automatically without any user interaction while edema localization requires user interaction for selection of the optimal segmentation result. The advantage of the method is in the use of unsupervised clustering which eliminates numerous

IV. Conclusions

A new method for automatic segmentation of CT head images of human spontaneous intracerebral brain hemorrhage has been proposed in this paper. The method has a two-level hierarchical structure and is based on unsupervised fuzzy clustering and image labeling techniques. The proposed method has been tested on real CT images and has performed correct segmentation. Future improvements of the method will include an expert system for higher segmentation accuracy and speed improvement. References

[1] G. I. Chiou and Jeng-Neng Hwang, \A neural network-based stochastic active model (NNS-SNAKE) for contour nding of distinct features", IEEE Transactions on Image Processing, vol. 4, pp. 1407{1416, 1995. [2] M. Ozkan, B. Dawant, and R. J. Maciunas, \Neural-networkbased segmentation of multi-modal medical images: A comparative and prospective study", IEEE Transactions on Medical Imaging, vol. 12, pp. 534{544, 1993. [3] S. Loncaric, A. P. Dhawan, T. Brott, and J. Broderick, \3-D image analysis of intracerebral brain hemorrhage", Computer Methods and Programs in Biomedicine, vol. 46, pp. 207{216, 1995. [4] J. G. Thomas, R. A. Petters, and P. Jeanty, \Automatic segmentation of ultrasound images using morphological operators", IEEE Transactions on Medical Imaging, vol. 10, pp. 180{185, 1991. [5] C. Li, D. B. Goldgof, and L. O. Hall, \Knowledge-based classi cation and tissue labeling of mr images of human brain", IEEE Transactions on Medical Imaging, vol. 12, pp. 740{750, 1993. [6] S. P. Raya, \Low-level segmentation of 3-d magnetic resonance brain images{a rule based system", IEEE Transactions on Medical Imaging, vol. 9, pp. 327{337, 1990. [7] S. Loncaric, D. Cosic, and A. Dhawan, \Segmentation of CT head images", in Proceedings of the 10th International Symposium Computer Assisted Radiology. 1996, Elsevier, Amsterdam. [8] L. O. Hall, A. M. Bensaid, L. P. Clarke, R. P. Velthuizen, M. S. Silbiger, and J. C. Bezdek, \A comparisonof neural network and fuzzy clustering techniques in segmenting magnetic resonance images of the brain", IEEE Transactions on Neural Networks, vol. 3, pp. 672{682, 1992. [9] I. Gath and A. Geva, \Unsupervised optimal fuzzy clustering", IEEE Transactions on PAMI, vol. 11, pp. 773{781, 1989. [10] R. Haralick and L. Shapiro, Computer and Robot Vision, vol. 2, Addison Wesley, 1992.

Suggest Documents