DISCUSSION AND CONCLUSIONS In this paper, we have presented an automated technique to segment transaxial data in 3D cardiac SPECT studies.The proposed methodology incorporates a priori knowledge of the LV structure and selectively employs various image processing and computer vision techniques to extract voxels corresponding to the left ventricle from the 3D dataset. This segmentation approach is an inherent part of a system to automatically determine the 3D orientation of the LV. Information regarding the pose of the LV is essential for obliquely slicing the 3D dataset to accurately quantify and visualize the blood perfusion in the myocardium. The volume segmentation methodology discussed here has been tested and evaluated in conjunction with the 3D LV orientation determination process. Since the presence of any other objects (in terms of voxels) besides the LV interferes with the pose determination process, accurate computation of the LV pose correlates well with correct object segmentation. Fifty(50) Thallium-201 and fifty(50) Tc-99m cardiac SPECT studies were used to evaluate the segmentation methodology. A qualitative analysis of the segmentation technique using these 100 datasets has yielded quite encouraging results. The ability to accurately determine the orientation of the LV [10] without human intervention quantitatively corroborates the correctness of this segmentation approach. ACKNOWLEDGMENTS This work was supported in part by grants R 29 LM04692 from the National Library of Medicine and 1 RO1 HL42052-01 from the National Institutes of Health. REFERENCES 1. Raya SP, “Rule based segmentation of multi-dimensional images,” Proceedings of the 10th annual NCGA conference, vol. 1, pp. 193-8, 1989. 2. Dellepiane S, Serpico SB, Vernazza G, Bruzzone S, and Regazzoni C, “An expert system for the recognition of anatomical organs from 3D database,” Proceedings of the Ninth Annual Conference of the IEEE Engineering in Medicine and Biology Society, vol. 2, pp. 933-4, 1987. 3. Kennedy DN, Filipek PA, and Caviness CJr., “Anatomic segmentation and volumetric calculations in nuclear magnetic resonance imaging,”, IEEE Transaction on Medical Imaging, vol. 8, no. 1, pp. 1-7, 1989. 4. Dellepiane S, Serpico SB, Vernazza L, and Vernazza G, “Structural analysis in medical imaging,” Proceedings of the 7th European Conference on Electrotechnics: Advanced Technologies and Processes in Communication and Power Systems - EUROCON 86, pp 522-9, April 1986. 5. Mortelmans L, Nuyts J, Van Pamel G, Van den Maegdenbergh V, De Roo M, and Suet, “A new thresholding method for volume determination by SPECT,” European Journal of Nuclear Medicine, vol. 12, no. 5-6, pp. 284-90, 1986. 6. Murase K, Tanada S, Yasuhara Y, Mogami H, Iio A, and Hamamoto K, “SPECT volume measurement using an automatic threshold selection method combined with a V filter,” European Journal of Nuclear Medicine, vol. 15, no. 1, pp. 21-5, 1989. 7. King MA, Long DT, and Brill AB, “SPECT volume quantification: influence of spacial resolution, source size and shape, and voxel size,” Medical Physics Journal, vol. 18, no. 5, pp 1016-24, 1991. 8. Long DT, King MA, and Sheehan, “Comparative evaluation of image segmentation methods for volume quantitation in SPECT,” Medical Physics Journal, vol. 19, no. 2, pp. 483-9, 1992. 9. Jain AK, “Fundamentals of Digital Image Processing,” Englewood Cliffs, N.J., Prentice Hall, 1989 (Chapter 9). 10. Mullick R, Ezquerra NF, Garcia EV, Cooke CD, and Folks RD, “3D Visualization of pose determination: Application to SPECT imaging,” Proceedings of the 1992 conference on Visualization in Biomedical Computing, SPIE vol. 1808, pp. 52-62, 1992. METHODOLOGY AND RESULTS The algorithm currently under development consists of a sequence of tasks performed either sequentially on each slice of the 3D data or in a unified way to the entire data volume. The various steps in the algorithm are briefly outlined here. The first step in the processing involves intelligent adjustment of the range of intensity values in the volume. This step is followed by appropriate scaling of intensity values in the dataset to lie within the determined intensity range. Before the GLH is computed, the range of slices for the histogram, is determined. This range is determined by computing the energy in each slice of the 3D dataset. If the slice with the maximum energy lies in the latter slices of the volume, suggesting the presence of the liver in the data, the latter slices are excluded when computing the GLH. Once the histogram is computed, the entries of the histogram are passed through a low-pass-filter to eliminate any possible noise in the statistics. The values in this “smoothed” histogram are then fit to a decaying exponential as shown in Figure 2. The exponential fit to the histogram is then employed to compute the optimal threshold value for the 3D data. The optimal threshold value is chosen to be that intensity value at which the slope of the exponential fit approaches zero. This computed threshold value is then restrained within predetermined limits based on the isotope used for the SPECT acquisition. The contents of each transaxial image are then processed using the morphological technique [9] of erosion using a 2x2 structuring element. The erosion process disconnects and thins out the walls of myocardium making it feasible to separate the two ventricles (left and right). A labeling technique is next employed to identify and mark each distinct mass in each image. All smaller masses except the largest one and those lying within a Fit pre-determined region (based on analysis of the mid-ventricular slice and Histogram in general the cardiac anatomy) in each image are deleted. At this stage all remaining masses are taken to be part of the left ventricular myocardium. Threshold Each segmented slice is then dilated to recover any possible viable data voxels. 3D connectivity of these 2D masses is also analyzed and the single largest connected mass in the 3D volume dataset is thus extracted. Figure 2 The final stages of this volume segmentation algorithm further excludes undesirable regions. This stage of processing is necessitated by the presence of voxels pertaining to the liver and other organs in the vicinity of the heart in a cardiac SPECT acquisition. This is achieved by limiting the radial expanse of the segmented data in each slice and by incrementally increasing the threshold value of each transaxial slice until the cavity (region of low voxel intensity) is visible in the mid-ventricular slice. This final step concludes the segmentation algorithm and results in a 3D dataset including only those voxels that are part of the left ventricular myocardium. The result of the application of this algorithm on the 3D dataset in Figure 1 is illustrated in Figure 3. It must be emphasized here that the entire process was carried out without any human intervention. The applications of this algorithm are many, especially in the realm of SPECT data analysis. The authors of this paper are currently employing the results of this segmentation algorithm for automatic determination of the left ventricle (LV) orientation and medial axis [10]. Figure 3: Result of automatic segmentation of 3D data in Figure 1. AUTOMATIC SEGMENTATION OF 3D CARDIAC SPECT IMAGERY Rakesh Mullick and Norberto F. Ezquerra* Dept. of Electrical Engineering and *College of Computing GVU Center, Georgia Institute of Technology, Atlanta, GA 30332 Email: rakesh@cc.gatech.edu ABSTRACT The automatic visualization and quantitative analysis of cardiac SPECT data requires the ability to automatically segment and extract voxels representing the heart. The attributes of the 3D data make this task quite challenging. In this paper, we attempt to address these issues and propose an algorithm which successfully detects the voxels belonging to the Left Ventricle (LV) of the heart and filters out the noise and all other interfering organs. The algorithm relies on various image processing and pattern analysis techniques as well as the constraints put forward by the anatomy. The final outcome of this algorithm is a segmented 3D dataset containing voxels pertaining only to the LV. This filtered dataset is then employed for automatic determination of LV orientation. The results show that this methodology is a very promising approach to segmentation of cardiac SPECT imagery. INTRODUCTION Significant work in the area of segmentation of medical imagery has been limited to high resolution magnetic resonance images [1, 2, 3, 4]. Some of these algorithms also employ techniques based on expert systems, neural networks and other high level image understanding systems. Research in the area of segmentation of SPECT data in particular, has been directed towards accurate volume determination of organs [5, 6, 7]. Various techniques [5, 6] for optimum segmentation based on a gray level histogram (GLH) and a V filter have been suggested in the literature. A comparative study of the image segmentation methods for volume quantification in SPECT [8] by Long et. al implies that a method based on 3D edge detection is most suitable for minimal operator intervention, accuracy, and consistency in estimation of object volume. In this paper, we present a methodology which consists of an algorithm unifying various image processing and computer vision techniques. In addition, more recent techniques of morphological image processing and connected component labeling have also been employed to further enhance the segmentation process [9]. DATA CHARACTERISTICS Inherent limitations of SPECT imaging, yields a challenging segmentation problem. The specific aim in nuclear medicine, and in particular Tl201 and Tc99m cardiac SPECT imagery is to study the physiology and not the structure of the imaged organ. However, it is desirable to infer structural information from these images. Thus, the functional information can be misleading, especially in the presence of hypoperfused regions in the data. This can make the interpretive task increasingly difficult. Figure 1: Transaxial slices from a Tc-99m cardiac SPECT acquisition. The constant “beating” of the heart has an additional blurring effect on this inherently low-resolution imaging modality. Current resolutions of such imaging systems range from 4 mm to 7 mm, resulting in small sized 3D datasets. Thus small errors (≤ 1 voxel) in localization of extracted features can lead to considerable anomalies in the quantitative analysis of the LV [7]. In addition, Compton scattering and other physical phenomena make the data quite noisy. Therefore, the data can be viewed as low-resolution, noisy, temporally integrated, and sparse. Therein lies the importance of proper segmentation, to minimize possible data loss and maximize the available data. A sample normal patient volume (3D) dataset in the form of consecutive transaxial slices is presented in Figure 1.