You are on page 1of 9

JDMS

24:329-336

November/December 2008

329

RESEARCH ARTICLES Image Segmentation for Evaluating Axillary Lymph Nodes


KEVIN D. EVANS, PHD, RT, RDMS, RVS, FSDMS STEFFEN SAMMET, MD, PHD YVETTE RAMOS, RDMS, RVT MICHAEL V. KNOPP, MD, PHD

A review is provided of the literature that has been published on image segmentation relative to sonography. Manual and automatic techniques for partitioning a sonogram are highlighted. In addition, a preliminary set of results is provided on the interrater reliability of the manual segmentation of axillary lymph nodes that have been sonographically imaged. A correlation between sonographers conducting manual segmentation is very high (r = 0.9 with P < .00 at the .01 alpha level). This work is set to provide additional information on lymph node cubic volume and the agreement between manual and automatic segmentation of axillary lymph nodes. Key words: image segmentation, breast sonography, axillary lymph nodes

From The Ohio State University, School of Allied Medical Professions, Columbus, Ohio. Correspondence: Kevin D. Evans, PhD, RT, RDMS, RVS, FSDMS, The Ohio State University, School of Allied Medical Professions, 410 W. 10th Avenue, 340 A. Atwell Hall, Columbus, OH 43210. E-mail: Evans.36@osu.edu. The authors thank GE Healthcare Ultrasound for providing the equipment, transducers, and software to conduct this research. DOI: 10.1177/8756479308324954

The use of computer algorithms for the delineation of anatomical structures or other regions of interest (ROIs) is called image segmentation.1 Image segmentation is not a new evaluation technique, but the use has proliferated to provide more information that is locked inside a medical image. Segmenting a medical image can be extremely helpful in quantifying tissue volumes, diagnosis, localization of pathology, and the study of anatomical structures.25 Unfortunately, image segmentation has been conducted with only a limited number of sonographic cases. Examples of successful clinical research using image segmentation of sonography images are those determining heel density, ovarian cysts, breast cysts, and fetal, liver, and cardiac pathology.615 The ability to segment anatomic and pathologic information from an image is only possible because of the composition of voxels. Extracting

330

JOURNAL OF DIAGNOSTIC MEDICAL SONOGRAPHY

November/December 2008

VOL. 24, NO. 6

information from within the image is done by partitioning the image into nonoverlapping, homogeneous regions based on the intensity of reflected acoustic energy. The clinical application of image segmentation has yielded a number of more effective techniques for conducting this process with sonograms. The following image segmentation techniques are commonly used with sonograms: textural classifiers, manual interaction to place ROIs, and deformable models.1 A brief description is provided of the image segmentation techniques that were considered for this feasibility study.

Literature Review
Mammography is an imaging modality in which segmentation techniques have been used extensively. In this area of research, a combination of multiple image segmentation techniques has been evaluated. An early study that was conducted with 2D digital mammography used segmentation of breast tumors in two steps.19 In this study, regions of interest were first extracted after a histogram was created, based on the frequency of pixel magnitude. The pixel frequency spanned a range of intensities from 70 to 120. Those pixel magnitudes that were greater than the threshold were deemed to be the breast tissue. After breast tissue was isolated, a modified Markov random field model was used to separate the tumor based on the extremely high pixel value that exceeded the typical breast tissue pixel density. Markov random field (MRF) models are considered advanced segmentation techniques, and this low-level model is based on the individual behavior of pixels relative to one another in the matrix.20 MRF models are classified as a form of clustering.21 These sets of segmentation steps were able to successfully isolate minimal breast cancers that were <10 mm. Another landmark study in this area was conducted to extract morphologic features of malignant and benign breast masses. This study used a combination of clustering, a deformable model, and a hybrid technique called speculation detection stages.22 One of the problems with this study was the need to digitize the film to begin the process of extraction. This allows the possibility of noise in the image and degrades the ability to discretely assess the pixels within the selected region. Breast imaging has embraced the use of computer-aided detection (CAD) software specifically developed for digital mammography. CAD makes use of layers of analysis that are focused on specific ROIs and classifies them for potential pathology. CADs primary function is to reduce the false-positive diagnoses that are made during breast imaging. CAD makes use of image analysis techniques that aid in making a diagnosis. The first

Segmentation Methods
1. Textural classifiers Clustering: Clustering is a form of textural classifying that depends on training data. In essence, the algorithm trains itself during the data analysis process.1 Once the region on the image is selected for extraction, the pixels within the extracted area are evaluated for a mean intensity, and subsequent segmentations are made based on finding like mean intensities.16 2. Regions of interest Region growing: Region growing requires that a manual interaction be made with the image and a seed placed in the area of interest. The algorithm moves outward from the seed to encompass voxels that are of like intensity and extracts them for the background image.1 3. Deformable models: Deformable models are a graphic representation of elastic theory that permits graphic molding around a selected voxel composition and deforms by responding to 17 energy internally. Deformable models can be manual or automated. Obviously, the reliability and consistency of automated deformable models are ideal to reduce bias and time.17 The snake is the most popular deformable model that has been used for image segmentation.18 Level-set segmentation allows a 3D image volume set to be extracted using a deformable surface model and adjusts throughout the volume.

IMAGE SEGMENTATION FOR EVALUATING AXILLARY LYMPH NODES / Evans et al

331

layer of analysis is contrast enhancement, and the next progressive layer is optimal image segmentation.23 In studies to develop optimum segmentation methods for CAD, an unsupervised Gaussian mixture method model and a hidden MRF model developed by Zang et al.24 were used. These analytical models were fused with a template of model images that represent normal pixel densities. The template model was considered a supervised test that was used to find the optimum segmentation maneuver to accentuate the CAD diagnosis. The study used 400 digital mammograms to determine the best mix of segmentation strategies with the CAD program. The result of the experiment demonstrated that a supervised segmentation method, as developed by Zang et al.,24 was a superior technique for CAD.23 This study was based on 2D images and building a system that was sensitive to gray-scale pixel differences. Calcifications in breast tissue also have been explored as a target for image segmentation and enhancement during the use of CAD. An experiment was conducted using 260 single-view film screen mammograms that were digitized and ready for CAD. In addition, image segmentation was performed using symmlet wavelet and a new technique called the donut filter.25 ROIs were placed over the 138 benign calcifications and 122 microcalcifications, and the signals that emanated from these particular pixels were used to cue segmentation. The study provided promising results and should be replicated with images from a variety of digitizers as well as digital mammography. The researchers also felt that other clinical applications could be found for this type of 2D image segmentation. Level-set segmentation, a specific method used with 3D volume image sets, can be affected by inherent noise from the imaging technique. The problem with noisy images within a 3D volume set is that smoothing or enhancing the data causes distortion of the anatomy, which has been targeted for segmentation. The trick is to find a method that strikes a balance between 3D resolution and inherent digital noise. The use of level-set segmentation relies on a surface-fitting strategy that helps in dealing with small-scale noise from the modality

and smoothes the intensity of the signal fluctuations in the volume data.26 There are many deformable models for segmenting 3D volume data. Using an implicit model, which specifies the surface as a level set and allows for the variation descending through the volume, would be ideal.27 This allows the data set to be defined by the deformable surface and vary dynamically over time. An additional concern is that volume data sets are collected in the X, Y, and Z planes. The Z plane has the lowest resolution due to data being collected in the axial direction. The introduction of this kind of inherent noise can be the source of inaccuracy within the targeted anatomy for segmentation. A group of researchers dealt with this problem by using multiple sets of scans on an object and merging them prior to segmentation. This hybrid image volume set was composed of individually higher resolution planes of information. The use of level-set segmentation on a hybrid 3D image volume was conducted using a magnetic resonance (MR) scan of a mouse embryo, a laser scan of a figurine, and an MR scan of a zucchini.26 The result was a high-resolution level-set model that contained all the significant features of the original scan for all three items. Because sonography is traditionally considered an imaging modality fraught with a certain level of digital noise, this segmentation technique could be quite helpful. Work conducted in Japan with image segmentation used 3D volume sets with the intent to better plan breast cancer surgeries.28,29 The 3D images were created from volume-rendered 2D sonograms and manually selected ROIs chosen from a histogram of pixel intensities. These 3D models were created while the patient was prepared for surgery. Creating these segmented models of the tumor helped to direct the surgeon to minimize the amount of tissue removed while maintaining appropriate borders. This work demonstrated the translational value of working with segmentation of 3D breast images and the potential to minimize the invasiveness of surgery. Previous research that has been conducted with the segmentation of sonographic images has demonstrated the value of increasing our diagnostic

332

JOURNAL OF DIAGNOSTIC MEDICAL SONOGRAPHY

November/December 2008

VOL. 24, NO. 6

capability and assisting the radiologist with additional diagnostic data. The evaluation of axillary lymph nodes has yet to be demonstrated with this set of imaging techniques, and the feasibility of creating these images remains to be explored. To fill this gap in the literature, we devised a feasibility study to explore the use of a manual technique for segmenting the sonographic voxel density of axillary lymph nodes.

Materials and Methods


Image segmentation has been used more extensively with computed tomography (CT), magnetic resonance imaging (MRI), and now digital mammography. These techniques have developed in response to the increasing ability to capture large volume sets of imaging information. Data sets have varied anatomic and pathologic information that may need to be located and segmented away for the larger data set. A robust set of maneuvers is needed to determine the coordinates of the item that will be segmented away from the background image. Predominately, sonography images are captured in 2D frames that are composed of pixels. This can limit the ability to accurately capture an images boundaries for segmentation. A volume set of image data, which is provided with 3D sonographic imaging, provides more digital information but has added complexities for conducting accurate image segmentation. Presently, the focus of this research team has been an attempt to use both manual and automated deformable models (snakes and level sets) to determine the best method for conducting segmentation of axillary lymph nodes. Attempts have been made to perform automated image segmentation with an existing data set to provide either a pixel count or a voxel count. The secondary data segmentation with automated techniques was attempted on the following sample images:
a. b. c. d. A 2D lymph node with color Doppler imaging A 2D lymph node without color Doppler imaging A 3D lymph node with color Doppler imaging A 3D lymph node without color Doppler imaging

FIGURE 1. 3D axillary lymph node image with color Doppler that was problematic for automatic image segmentation.

In addition, a manual deformable model of image segmentation was used to provide a voxel count on a sample 3D lymph node image without color Doppler imaging (see Figure 1). Unfortunately, these attempts were not successful, in part, because these images were not collected for this expressed purpose. A fresh data collection was needed to properly determine the feasibility of these techniques to isolate lymph nodes from surrounding breast tissue.
PROTOCOL FOR SEGMENTATION

The proposal was to use normal female volunteers to allow for a collection of images of lymph nodes in the axilla. This allowed for a manual segmentation with a sonographic transducer that was specifically built to make volume measurements and segment the image at the time it was collected. The TRU 3D/Vocal transducer setup was provided by General Electric Healthcare along with a GE Logic 9 (Milwaukee, Wisconsin). This equipment allowed the segmentation to be done manually using the deformable model techniques. The Institutional Review Board of Ohio State University approved this project. The number of volunteers was set at 40 to achieve a moderate effect size of .5 at an alpha of .05 with a power of .59.

IMAGE SEGMENTATION FOR EVALUATING AXILLARY LYMPH NODES / Evans et al

333

FIGURE 2. Image segmentation of an axillary lymph node using the deformable model snake technique.

FIGURE 4. Small node that would be classified as abnormal and the model created with the snake technique.

FIGURE 3. Larger, more oval node segmented with a deformable model snake technique.

FIGURE 5. Image segmentation of a small node with image optimization prior to manual segmentation.

Each normal female volunteer consented to have 20 minutes of diagnostic medical sonography, and after three lymph nodes were imaged in volume sets, they were dismissed. The volume sets of lymph node images were then manually segmented using the deformable model snake (see Figures 25). Once the images had been through

the segmentation process, a cubic volume was determined, and the voxel density was determined mathematically. All the images were de-identified and retained on a flash drive that was kept in a locked office and used only for transferring the volume data sets to Dr. Sammets computer lab for future automated processing.

334

JOURNAL OF DIAGNOSTIC MEDICAL SONOGRAPHY

November/December 2008

VOL. 24, NO. 6

Results
Preliminary results are provided on 23 participants who provided a total of 45 lymph nodes for evaluation. The mean age of these 23 participants is 24 years of age, with the hope that younger participants would provide more normal-shaped nodes free of inflammatory or chronic disease changes. One of the key initiatives at this juncture in the research was to use these 45 lymph nodes for manual segmentation but to check for interrater reliability. The principal investigator (KE) and the application specialist sonographer (YR) worked independently to perform manual image segmentation for each node. The resulting 3D model that was created was recorded and the cubic volume compared for consistency. A mean cubic volume of .28 m3 (SD .26) was determined for the 45 nodes that were manually segmented by the principal investigator. The application specialist provided a mean cubic volume of .31 m3 (SD .27) for those nodes that she manually segmented. The Pearson correlation coefficient was chosen to determine the statistical correlation of measures that were made by the two sonographers per node. The set of nodes was compared between sonographers, according to the order in which the nodes were captured (node 1, node 2, and node 3). For node 1 in the series, the correlation coefficient was r = 0.91, which indicates a very high positive correlation between the sonographers measurements for all those first nodes segmented in the series. The statistical significance of this agreement was P < .00 at the .01 alpha level. The scatterplot for those manually segmented nodal volumes is depicted in Figure 6. For node 2 in the series, the correlation coefficient was r = 0.97, which indicates a very high positive correlation between the sonographers measurements for all those second nodes segmented in the series. The statistical significance of this agreement was P < .00 at the .01 alpha level. The scatterplot for those manually segmented nodal volumes is depicted in Figure 7. For node 3 in the series, the correlation coefficient was r = 0.99, which again indicates a very high positive correlation between the sonographers

measurements for all those third nodes segmented in the series. The statistical significance of this agreement was P < .00 at the .01 alpha level. The scatterplot for those manually segmented nodal volumes is depicted in Figure 8.

Discussion
The very high correlation between the sonographers measurements for manual segmentation of the axillary lymph nodes imaged is a promising

FIGURE 6. Scatterplot of the manual segmentation of the cubic volume of the first axillary lymph node imaged.

FIGURE 7. Scatterplot of the manual segmentation of the cubic volume of the second axillary lymph node imaged.

IMAGE SEGMENTATION FOR EVALUATING AXILLARY LYMPH NODES / Evans et al

335

sonographic breast lesions in a study of 400 sonograms that spanned a variety of pathologies.31 The researchers concluded that automated and manual segmentation were similar in their ability to consistently delineate borders of breast masses. The next phase of the current research will be to complete a comparison between automated and manual segmentation with the existing data set. The limitations inherent with a feasibility study of this size are that the results are unique and may not generalize to a larger population.

Conclusion
FIGURE 8. Scatterplot of the manual segmentation of the cubic volume of the third axillary lymph node imaged.

early result for the continuation of this research. One criticism of manual segmentation is that the variability of the models is attributed to the operator; however, this does not appear to be a factor in the research that we are conducting. The scatterplots are highly effective tools that can be used to look for those nodal volumes that do not line up close to 1.0. These may be nodes that need to be reviewed to determine what caused the sonographers to disagree on their method of manual segmentation of the cubic volume. The impact of this research will be boosted by securing all 40 volunteers and completing the manual segmentation of axillary lymph nodes that are visualized. Once these nodes are segmented, the data set will be segmented automatically using automated level sets or region-growing algorithms. Interestingly, a study was done using a mixed model of image segmentation in which sonographic images were partitioned using a method of outlining the contour of the image, with the automated regiongrowing technique completing the segmentation.30 The software was trained to be more accurate by incorporating the contour outlines provided by the operator. In this study, the gallbladder and kidney were the organs segmented on sample sonograms. In addition, segmentation with both automated and manual techniques was conducted with

Continued use of manual segmentation with sonographic images will help build a body of evidence-based research and promote translational implementation. The next specific aim is not only to demonstrate consistency with segmenting lymph nodes but also to compute the normal mean volume of a lymph node. This will allow the research team to compute the number of voxels that are expected to comprise a nodes normal cubic volume.

References
1. Pham D, Xu C, Prince J: A survey of current methods in medical image segmentation. Ann Rev Biomed Eng 2000;2:315337. 2. Larie S, Abukmeil S: Brain abnormality in schizophrenia: a systematic and quantitative review of magnetic resonance imaging studies. Br J Psychiarty 1998;172: 110120. 3. Taylor P: Invited review: computer aids for decisionmaking in diagnostic radiologyA literature review. Br J Radiol 1995;68:945957. 4. Zijdenbos A, Dawant B: Brain segmentation and white matter lesion detection in MRI images. Crit Rev Biomed Eng 1994;22:401465. 5. Worth A, Makris N, Caviness V, Kennedy D: Neuroanatomical segmentation of MRI: technological objectives. Int J Patt Recog Art Intell 1997;11:11611187. 6. Zimmerand Y, Tepper R, Akselrod S: A two-dimensional extension of minimum cross entropy thresholding for the segmentation of ultrasound images. Ultrasound Med Biol 1996;22:11831190. 7. Coppini G, Poli R, Valli G: Recovery of the 3-D shape of the left ventricle from echocardiographic images. IEEE Trans Med Imaging 1995;14:301317.

336

JOURNAL OF DIAGNOSTIC MEDICAL SONOGRAPHY

November/December 2008

VOL. 24, NO. 6

8. Sebbahi A, Herment A, deCessare A, Mousseaux E: Multimodality cardiovascular image segmentation using a deformable contour model. Comput Med Imaging Graphics 1997;21:7989. 9. Jones T, Metaxas D: Segmentation using deformable models with affinity-based localization. Lect Notes Comput Sci 1997;1205:5362. 10. Kucera D, Martin R: Segmentation of sequences of echocardiographic images using a simplified 3D active contour model with region based external forces. Comput Med Imaging Graphics 1997;21:121. 11. Lefebvre F, Berger G, Laugier P: Automatic detection of the boundary of the calcaneus from ultrasound parametric imaging using an active contour model: clinical assessment. IEEE Trans Med Imaging 1998;17:4552. 12. Caselles V, Catte F, Coll T, Dibos F: A geometric model for active contours. Numerishche Mathematik 1993;66: 131. 13. Pathak S, Chalana V, Kim Y: Interactive automatic fetal head measurements from ultrasound images using multimedia computer technology. Ultrasound Med Biol 1997;23:665673. 14. Yezzi A, Kichenassamy S, Kumar A, Oliver P, Tannenbaum A: A geometric snake model for segmentation of medical imagery. IEEE Trans Med Imaging 1997;16:199209. 15. Kadah Y, Farag A, Zurada A, Badawi A, Youssef A: Classification algorithms for quantitative tissue characteristics of diffuse liver disease from ultrasound images. IEEE Trans Med Imaging 1996;15:466478. 16. Jain A, Dubes R: Algorithms for Clustering Data. Englewood Cliffs, NJ: Prentice Hall; 1988. 17. McInerney T, Terzopoulos D: Deformable models in medical image analysis: a survey. Med Image Anal 1996; 1:91108. 18. Kass M, Witkin A, Terzopoulos D: Snakes: active contour models. Int J Comp Vision 1988;1:321331. 19. Kalleergi M, Clarke L, Jain V: Markov random field for tumor detection in digital mammography. IEEE Trans Med Imaging 1995;14:565576. 20. Farag A, Ahmed M, E-Baz A, Hassan H: Advanced segmentation techniques, in Suri J, Wilson D, Laxminarayan S (eds): Handbook of Biomedical Image Analysis: Segmentation Models: Part A. Vol. 1. New York, Kluwer Academic/Plenum, 2005, pp 479533.

21. Pappas T: An adaptive clustering algorithm for image segmentation. IEEE Trans Signal Processing 1995;40: 902914. 22. Sahiner B, Chan H, Petrick N, Helvie M, Hadjiiski L: Improvement of mammographic mass characterization using speculation measures and morphological features. Am Assoc Physicists Med 2001;28:14551465. 23. Singh S, Bovis K: A knowledge-based scheme for digital mammography, in Suri J, Wilson D, Laxminarayan S (eds): Handbook of Biomedical Image Analysis: Segmentation Models: Part B. Vol. 2. New York: Kluwer Academic/Plenum, 2005, pp 591660. 24. Zang Y, Brady M, Smith S: Segmentation of brain MR images through a hidden Markov random field model and the expectation minimization algorithm. IEEE Trans Med Imaging 2001;20:4557. 25. Kallergi M, Heine J, Tembey M: Computer-aided diagnosis of mammographic calcification clusters: impact of segmentation, in Suri J, Wilson D, Laxminarayan S (eds): Handbook of Biomedical Image Analysis: Segmentation Models: Part B. Vol. 2. New York: Kluwer Academic/ Plenum, 2005, pp 707751. 26. Breen D, Whitaker R, Museth K, Zhukov L: Level set segmentation of biological volume datasets, in Suri J, Wilson D, Laxminarayan S (eds): Handbook of Biomedical Image Analysis: Segmentation Models: Part A. Vol. 1. New York, Kluwer Academic/Plenum, 2005, pp 415533. 27. Osher S, Sethian J: Fronts propagating with curvature dependent speed: algorithms based on Hamilton-Jacobi formulations. J Comp Phys 1988;79:1249. 28. Sato Y, Nakamoto M, Tamaki Y, et al: Image guidance of breast cancer surgery using 3-D ultrasound images and augmented reality visualization. IEEE Trans Med Imaging 1998;17:681693. 29. Inoue T, Tamaki Y, Sato Y, et al: Three dimensional ultrasound imaging of breast cancer by a real-time intraoperative navigation system. Breast Cancer (Japan) 2005;12: 122129. 30. Levienaise-Obadia B, Gee A: Adaptive segmentation of ultrasound images. Image Vision Comp 1999;17: 583588. 31. Horsch K, Giger M, Venta L, Vyborny C: Automatic segmentation of breast lesions on ultrasound. Med Phys 2001;28:16521659.

JDMS

24:337-338

November/December 2008

337

SDMS-JDMS CME TEST


Article: Image Segmentation for Evaluating Axillary Lymph Nodes Authors: Kevin D. Evans, PhD, RT, RDMS, RVS, FSDMS Steffen Sammet, MD, PhD Yvette Ramos, RDMS, RVT Michael V. Knopp, MD, PhD Category: Other Credit: 1.0 CME Objectives: After studying the article titled Image Segmentation for Evaluating Axillary Lymph Nodes, you will be able to: 1. Describe how image segmentation can be used to partition a 3D digital image of axillary lymph nodes. 2. Discuss how image segmentation can be accomplished with either manual or automatic processes. 3. Define image segmentation. 1. Image segmentation is defined by which of the following? a. The slicing of images for making digital maps b. The use of computer algorithms for the delineation of anatomical structures c. The compression of images to load on the PACS to conserve memory d. The joining of images to create a 2D image 2. Image segmentation has been conducted in which of the following sonography examinations? a. To study cardiac pathology b. To determine gestational age based on the nasal bone c. Several solid breast masses d. The adult liver 3. Which of the following image segmentation processes makes use of the electronic snake? a. Textural classifiers b. Regions of interest c. Deformable models d. Clustering 4. What was the mean age of the female participants involved in this study? a. 40 years of age b. 20 years of age c. 65 years of age d. 24 years of age 5. What was the correlation coefficient of agreement between the two sonographers involved in manual image segmentation of axillary lymph nodes? a. 1.0 b. 0.5 c. 2.0 d. 0.9 6. The use of CAD is a form of image segmentation that used which of the following as an innovative filter in a 2005 research study? a. Wire mesh b. A digitizer c. A donut d. Electronic manipulation 7. Which of the following image sets were not used for automatic image segmentation? a. A 2D lymph node with power Doppler imaging b. A 2D lymph node without color Doppler imaging c. A 3D lymph node with color Doppler imaging d. A 3D lymph node without color Doppler imaging 8. Which image segmentation method uses a seed for region growing automatically in the image? a. Textural classifiers b. Regions of interest c. Deformable models d. Clustering 9. Which image segmentation method can be grossly affected by digital noise? a. Textural classifiers b. Regions of interest c. Deformable models d. Clustering 10. What was the mean cubic volume of the lymph nodes that were extracted in the research by Evans et al.? a. 10 cm3 b. 0.50 cm3 c. 0.28 cm3 d. 1.0 cm3

You might also like