CHAPTER 4 EXPERIMENTS & RESULTS
4.6 Comparison of the features
Table 4-2 lists the Mahalanobis distances (inter-distances) of the used features between the liver and the sponge. It is clear that the coarseness is a good visual feature to distinguish the echotexture of the fully speckles and that of the soft tissue. Contrast is probably a good feature too, but it suffers from the BSC.
Feature Coarseness Contrast Busyness Complexity
Without BSC 0.5880 0.5008 0.2805 0.0223
With BSC 0.6242 0.3737 0.3070 0.0294
Table 4-2
Table 4-3 lists the inter-distance of the used features between the normal liver and liver cirrhosis. It shows that the busyness is very good at discriminating the cirrhosis
0 2 4 6 8 10
0 0.05 0.1 0.15 0.2 0.25
cirrhosis normal sponge
Figure 32: Distribution of complexity values with the unification of sampling rate.
Feature Coarseness Contrast Busyness Complexity
Without BSC 0.3902 0.2448 0.8714 0.3236
With BSC 0.3157 0.4469 1.0984 0.4131
Table 4-3
Combining several features does not improve the separability by these features.
Figure 33 shows the combination of busyness and coarseness, the three classes cluster together. Even with the BSC, the combination is not effective (Figure 34).
6 8 10 12 14
0 20 40 60 80 100 120
coarseness
b u syn e ss
without BSC
sponge cirrhosis normal
Figure 33: Feature distributions without BSC. The selected features are busyness and coarseness.
60 80 100 120 140 0
1 2 3 4 5 6 7
with BSC
coarseness
bus y n es s
sponge cirrhosis normal
Figure 34: Feature distributions with the BSC. The selected features are busyness and coarseness.
Chapter 5
Discussion and Conclusion
DSC makes deep echotexture coarser than shallow echotexture is, and theoretically back-scan conversion would reduce the depth dependence. However, it is found that shallow echotexture is slightly coarser than the one at deep depth, according to results shown in Figure 13 and the mean coarseness value at different depths mentioned above (Figure 16). This may be due to the dynamic focusing and/or the diffraction effect, both of them can make the image resolution uneven in range.
Comparing Figure 17 and Figure 18, two important results can be found. The first is that the coarseness values of sponge images are distributed quite compactly no matter whether the preprocessing is applied or not. On the other hand, the variation of the coarseness values of liver images is reduced by preprocessing; this is the main reason that the separability between sponge and liver images is improved by the preprocessing.
The second is that the variations of the coarseness values of sponges are much smaller than that of liver images. This is due to the fact that the liver cells are organized in lobules by collagen, and collagen is a strongly scattering medium which leads to the structural scattering in the liver images and makes the liver echotexture inhomogeneous.
This result shows that the coarseness measure is more effective for the homogeneous sponge images than for inhomogeneous liver images.
Some interesting results can be found in this study. Firstly, in both Figure 25 and Figure 26, variation of coarseness values of normal liver are all smaller than that of the cirrhotic case. This means that the texture of normal liver is more homogeneous than that of cirrhotic livers. Especially in Figure 26 the distributions of coarseness values of texture of normal liver with preprocessing, its distribution is symmetric without long tail as in the case in Figure 25, this reflects the usual viewpoint that “ultrasound image texture of normal liver is homogeneous”; here, the homogeneity is confirmed via the help of preprocessing. Such a result reveals that the human vision especially the
“expert’s vision” may do his/her own compensation to reach the conclusion that “an unpreprocessed ultrasound image of normal liver is homogeneous”. However, this is not the case for computational vision. Secondly, the mean coarseness value of sponges in Figure 25 is 7.7872 ( fs−cos), while that of normal livers is about 8.7288 ( fn−cos), and cirrhosis is 9.3403 (fc−cos). These coarseness values can be transformed to be their
physical coarseness values in unit of length via the help of sponge image. Assuming that the sponge image is homogeneous, the dimension of speckle cell has a mean size d = s
0.35 mm calculated by the FWHM of ACVF. For fully developed speckle, its coarseness measure should represent its physical speckle size. Then, the physical coarseness of normal livers relative to sponge is
(
fn−cos fs−cos)
×ds=0.39 mm and that of cirrhotic livers is(
fc−cos fs−cos)
×ds=0.42 mm. This confirms the usual viewpointthat “ultrasound image texture of cirrhotic livers is coarser than that of normal liver.”
With consideration to the characteristics of ultrasonic wave and the histology of the liver, “coarseness” proposed by Amadasun may be treated as to measure the homogeneity in the ultrasonic resolution cell. In case that the concentration of scatterers is higher than a threshold and the scattering power of one scatterer is close to the others within a resolution cell, the scatterers may be regarded as spreading homogeneously over the resolution cell. On the contrary, if the resolution cell contains some particles which have larger scattering power than other scatterers, the resolution cell is inhomogeneous. Therefore, coarseness may distinguish the echotexture of fully developed speckle from that of the liver. It is difficult to separate liver cirrhosis from normal liver, because the composition of liver in a resolution cell is not homogeneous.
The parenchyma of spleen is more homogeneous than that of the liver, and consequently the coarseness feature may be used to separate the echotexture of the spleen from that of the liver.
In Table 4-2, the performance of the contrast feature to distinguish the liver and sponge diminishes with BSC, but it rises to separate the cirrhosis and normal liver with BSC in Table 4-3. It is also mentioned that the ROI of the liver without BSC contains more Class 2 scattering than that with BSC. Thus if the shape of the ROI without BSC is modified like a “fan” (the right in Figure 10), the performance of the contrast feature
to separate the echotexture of the liver and sponge would be better with BSC. However, because the separability of classifying the liver and sponge is lower than that of the coarseness feature, plus the low separability of distinguishing the normal liver and cirrhosis, the contrast feature is not a good feature for analyzing the echotexture of soft tissue.
Even though the busyness feature could not separate the liver and sponge, it is workable to discriminate the liver cirrhosis from normal liver because it suppresses the locally high contrast cause by collagen, but keep the speckles. That is, the busyness is to measure the cluttering caused by the fully developed speckles, and we may surmise the busyness feature may respond to the tissue structure in the liver.
High spatial correlation of the ultrasound echotexture means the low contrast in a neighborhood, leading to a smaller number of primitives with different intensity in the echotexture. The primitives are monotonous, and then the complexity feature proposed by Amadasun is useless for ultrasound image texture.
It is clear that speckles dominate in the texture of ultrasound image, and the ability of Amadasun’s measure is to estimate the characteristic of speckles for ultrasound image texture. Visually discriminating cross-linked reticulum from speckles is based on their scale and contrast, and BSC with sampling rate unification helps reduce the distortion caused by the non-uniform sampling rate. Speckle reduction would change
the characteristic of echotexture, such as the distribution of gray level and the Laplacian.
Thus speckle reduction would affect Amadasun’s measure of echotexture. Amadasun’s measure, like most other texture analysis, is surely instrument dependent. Coarseness of echotexture is one of many possible features for diagnosing liver fibrosis, such as shear modulus [45, 46]; we believe that no single feature might be used as a dominant index for clinical diagnosis. Therefore, Amadasun’s measure could be used with other features to reach a better diagnosis for liver cirrhosis.
In this study, we propose an approach to quantify the visual properties of ultrasound image texture. As these features are comparative respectively, ultrasound images of sponge may be used as the references. It is demonstrated that Amadasun’s measure is useful to estimate the characteristics of the ultrasound image texture.
Furthermore, with the unification of sampling format and rate in the axial and lateral directions to reduce the distortion caused by scanning format, Amadasun’s measure is very effective in estimating the coarseness of fully developed speckle texture. In spite of this effectiveness, the use of “coarseness” does not discriminate the images of cirrhotic liver from that of normal livers well. This is due to the reason that the “coarseness” is not so effective in estimating the coarseness of inhomogeneous textures. The
“busyness” feature is good at separating the echotexture of the cirrhosis and normal liver; it may respond the distribution of connective tissue in the ultrasound liver image.
In addition to the quantification of ultrasound echotexture and the unification of the sampling format and rate for the analysis of B-mode image, the most contribution of this study is to combine the histology of the tissue and the ultrasound echotexture.
Reference
[1] B. S. Garra, M. F. Insana, T. H. Shawker, R. F. Wagner, M. Bradford, and M.
Russell, "Quantitative ultrasonic detection and classification of diffuse liver disease. Comparison with human observer performance," Investigative Radiology, vol. 24, pp. 196-203, Mar 1989.
[2] P.-M. Yang, G.-T. Huang, J. T. Lin, J.-C. Sheu, M.-Y. Lai, I. J. Su, H. C. Hsu, D.-S. Chen, T.-H. Wang, and J. L. Sung, "Ultrasonography in the diagnosis of benign diffuse parenchymal liver diseases: a prospective study," Journal of the Formosan Medical Association, vol. 87, pp. 966-977, Oct 1988.
[3] T. Nishiura, H. Watanabe, M. Ito, Y. Matsuoka, K. Yano, M. Daikoku, H.
Yatsuhashi, K. Dohmen, and H. Ishibashi, "Ultrasound evaluation of the fibrosis stage in chronic liver disease by the simultaneous use of low and high frequency probes," British Journal of Radiology, vol. 78, pp. 189-197, Mar 2005.
[4] C.-M. Wu, Y.-C. Chen, and K. S. Hsieh, "Texture Features for Classification of Ultrasonic Liver Images," IEEE Transactions on Medical Imaging, vol. 11, pp.
141-152, Jun 1992.
[5] S. Tanaka, O. Oshikawa, T. Sasaki, T. Ioka, and H. Tsukuma, "Evaluation of tissue harmonic imaging for the diagnosis of focal liver lesions," Ultrasound in Medicine and Biology, vol. 26, pp. 183-187, Feb 2000.
[6] M.-H. Horng, Y.-N. Sun, and X.-Z. Lin, "Texture feature coding method for classification of liver sonography," Computerized Medical Imaging and Graphics, vol. 26, pp. 33-42, Jan-Feb 2002.
[7] P. W. Ralls, M. B. Johnson, G. C. Kanel, D. M. Dobalian, M. Colletti, W. D. J.
Boswell, D. R. Radin, and J. M. Halls, "FM sonography in diffuse liver disease:
prospective assessment and blinded analysis," Radiology, vol. 161, pp. 451-454, 1986.
[8] H. Fukuda, M. Ebara, A. Kobayashi, N. Sugiura, M. Yoshikawa, H. Saisho, K.
Kato, F. Kondo, and T. Yahagi, "Irregularity of parenchymal echo patterns of liver analyzed with a neural network and risk of hepatocellular carcinoma in liver cirrhosis," Oncology, vol. 63, pp. 270-279, 2002.
[9] H. Popper, "Pathologic aspects of cirrhosis: A review," The American journal of pathology, vol. 87, pp. 228-264, 1977.
[10] E. Lewis, "Screening for diffuse and focal liver disease: the case for hepatic sonography," Journal of Clinical Ultrasound, vol. 12, pp. 67-73, 1984.
[11] U. Raeth, D. Schlaps, B. Limberg, I. Zuna, A. Lorenz, G. van Kaick, W. J.
Lorenz, and B. Kommerell, "Diagnostic accuracy of computerized B-scan texture analysis and conventional ultrasonography in diffuse parenchymal and malignant liver disease," Journal of Clinical Ultrasound, vol. 13, pp. 87-99, 1985.
[12] R. M. Haralick, "Statistical and structural approaches to texture," Proceedings of the IEEE, vol. 67, pp. 786-804, 1979.
[13] T. Kimura, M. Ebara, M. Ohto, and F. Kondo, "Classification of liver cirrhosis based on parenchymal echo patterns and its clinical usefulness for diagnosis of liver cirrhosis," Nippon Shokakibyo Gakkai Zasshi, vol. 86, pp. 1473-1485, 1989.
[14] H. Fukuda, M. Ebara, A. Kobayashi, N. Sugiura, M. Yoshikawa, H. Saisho, F.
Kondo, S. Yoshino, and T. Yahagi, "An image analyzing system using an artificial neural network for evaluating the parenchymal echo pattern of cirrhotic liver and chronic hepatitis," IEEE Transactions on Biomedical Engineering, vol.
45, pp. 396-400, Mar 1998.
[15] J. T. M. Verhoeven and J. M. Thijssen, "Potential of Fractal Analysis for Lesion Detection in Echographic Images," Ultrasonic Imaging, vol. 15, pp. 304-323, 1993.
[16] Y. Wun and R. Chung, "Ultrasound characterization by stable statistical patterns," Computer Methods and Programs in Biomedicine, vol. 55, pp.
117-126, 1998.
[17] R. Momenan, R. F. Wagner, B. S. Garra, M. H. Loew, and M. F. Insana, "Image staining and differential diagnosis of ultrasound scans based on the Mahalanobis distance," IEEE Transactions on Medical Imaging, vol. 13, pp. 37-47, Mar 1994.
[18] E. Caturelli, L. Castellano, S. Fusilli, B. Palmentieri, G. A. Niro, C. del Vecchio-Blanco, A. Andriulli, and I. de Sio, "Coarse nodular US pattern in hepatic cirrhosis: Risk for hepatocellular carcinoma," Radiology, vol. 226, pp.
691-697, Mar 2003.
echograms: 3-D simulations and experiments of the effects of diffraction and scatterer density," Ultrasonic Imaging, vol. 7, pp. 142-160, 1985.
[20] F. M. J. Valckx, J. M. Thijssen, A. J. van Geemen, J. J. Rotteveel, and R. A.
Mullaart, "Calibrated Parametric Medical Ultrasound Imaging," Ultrasonic Imaging, vol. 22, pp. 57-72, 2000.
[21] S. L. Friedman, "Liver fibrosis – from bench to bedside," Journal of Hepatology, vol. 38, pp. S38-53, 2003.
[22] D. C. Rockey and D. M. Bissell, "Noninvasive measures of liver fibrosis,"
Hepatology, vol. 43, pp. S113-S120, Feb 2006.
[23] T. Szabo, Diagnostic Ultrasound Imaging: Inside Out: Academic Press, 2004.
[24] J. M. Thijssen, "Ultrasonic speckle formation, analysis and processing applied to tissue characterization," Pattern Recognition Letters, vol. 24, pp. 659-675, 2003.
[25] D. Nicholas, D. Nassiri, P. Garbutt, and C. Hill, "Tissue characterization from ultrasound B-scan data," Ultrasound in Medicine and Biology, vol. 12, pp.
135-143, 1986.
[26] I. Akiyama, T. Saito, M. Nakamura, N. Taniguchi, and K. Itoh, "Tissue characterization by using fractal dimension of B-scan image," in Proc. IEEE Ultrasonics Symposium, 1990, pp. 1353-1355, 1990.
[27] B. J. Oosterveld, J. M. Thijssen, P. C. Hartman, and G. J. E. Rosenbusch,
"Detection of diffuse liver disease by quantitative echography:dependence on a priori choice of parameters," Ultrasound in Medicine and Biology, vol. 19, pp.
21-25, 1993.
[28] F. S. Cohen and G. Georgiou, "Detecting and estimating structure regularity of soft tissue organs from ultrasound images," in IEEE International Conference on Image Processing, pp. 488-491, 1995.
[29] A. Mojsilovic, M. V. Popovic, and D. Sevic, "Classification of the ultrasound liver images with the 2N×1-D wavelet transform," in IEEE International Conference on Image Processing, pp. 367-370, 1996.
[30] H. J. Huisman, J. M. Thijssen, T. Wagener, and G. J. E. Rosenbusch,
"Quantitative Ultrasonic Analysis of Liver Metastases," Ultrasound in Medicine and Biology, vol. 24, pp. 67-77, 1998.
[31] W.-C. Yeh, S.-W. Huang, and P.-C. Li, "Liver fibrosis grade classification with B-mode ultrasound," Ultrasound in Medicine and Biology, vol. 29, pp.
1229-1235, 2003.
[32] A. Rosenfeld and M. Thurston, "Edge and Curve Detection for Visual Scene Analysis," IEEE Transactions on Computers, vol. C-20, pp. 562-569, 1971.
[33] J. S. Weszka, C. R. Dyer, and A. Rosenfeld, "A comparative study of texture measures for terrain classification," IEEE Transactions on Systems, Man, and Cybernetics, vol. 6, pp. 269-285, 1976.
[34] M. Amadasun and R. King, "Textural features corresponding to textural properties," IEEE Transactions on Systems, Man, and Cybernetics, vol. 19, pp.
1264-1274, 1989.
[35] R. M. Haralick, K. Shanmugam, and I. Dinstein, "Textural features for image classification," IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. SMC-3, pp. 610-621, 1973.
[36] C. Lai, J. Tsao, and M. Lo, "The effects of sampling rate on the texture separability of ultrasound images," in Conf Proc. IEEE Engineering in Medicine and Biology Society, pp. 4811-4814, 2006.
[37] I. I. Matalka, O. M. Al-Jarrah, and T. M. Manasrah, "Quantitative assessment of liver fibrosis: a novel automated image analysis method," Liver International, vol. 26, pp. 1054-1064, Nov 2006.
[38] I. Boniatis, L. Costaridou, D. Cavouras, I. Kalatzis, E. Panagiotopoulos, and G.
Panayiotakis, "Assessing hip osteoarthritis severity utilizing a probabilistic neural network based classification scheme," Medical Engineering & Physics, vol. 29, pp. 227-237, 2007.
[39] C. Castella, K. Kinkel, M. P. Eckstein, P. E. Sottas, F. R. Verdun, and F. O.
Bochud, "Semiautomatic mammographic parenchymal patterns classification using multiple statistical features," Academic Radiology, vol. 14, pp. 1486-1499, Dec 2007.
[40] F. W. Campbell and J. G. Robson, "Application of Fourier Analysis to Visibility of Gratings," Journal of Physiology, vol. 197, pp. 551-566, 1968.
[41] C. Fritsch, M. Parrilla, O. Martınez, and D. Jimenez, "A multirate scan
[42] A. P. Berkhoff, H. J. Huisman, J. M. Thijssen, E. M. G. P. Jacobs, and R. J. F.
Homan, "Fast Scan Conversion Algorithms for Displaying Ultrasound Sector Images," Ultrasonic Imaging, vol. 16, pp. 87-108, 1994.
[43] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal Processing, 2nd ed. London: Prentice Hall, 1999.
[44] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification, 2 ed. New York:
Wiley-Interscience, 2000.
Appendix
Neighborhood Gray-Tone Difference Matrix
Amadasun’s measure is based on a vector s(i), called neighborhood gray-tone difference matrix (NGTDM). s(i) is a column vector in which every element is
calculated corresponding to the gray level i in the image. Its entries are computed based on measuring the difference between the gray level of a pixel and the average gray level computed over a square, sliding window centered at the pixel. The original definition of NGTDM from Amadasun is ROI-size dependent; to make feature of ROI with different size comparable and rational, the definition of NGTDM must be modified as follows.
Suppose the image value fi( lk, ) at pixel(k,l) is i, i = 0, 1, …, G . Let h Aikl be the
normalization factor for different size of ROI. Amadasun’s experiments showed that the
neighborhood size being 1 or 2 does not affect the performance of the proposed features much, since this measure is based on the concept of patches with low local contrast. For that reason, d is set to be one for reducing computation complexity.