A new approach to ultrasonic liver image classification
全文
(2) IEICE TRANS. INF. & SYST., VOL.E83–D, NO.6 JUNE 2000. 1302. (BPNN) in our method because of its inherent shortcomings such as the expensive training process, susceptibility to the local minimum, difficulty in adding new training samples et al. The Probabilistic Neural Network (PNN) [12] can avoid these shortcomings and approach the Bayes optimal, thus it is employed as a classifier in our method. Sixty-six ultrasonic liver samples are used to test the performance of the proposed method. The experimental results show that our method can produce about 88% correct classification rate using four features. This result is superior to the individual use of the co-occurrence matrices, the Fourier power spectrum, and the texture spectrum measures [1]–[5]. In comparison with the MTDV approach [9], which used ten texture features for sonographic evaluation, our method takes only four features to achieve the analogous classification rate for an even tougher classification task that is the progressive assessment of the three diffuse liver states. This implies that the selected features by the proposed method are more abundant in texture information for liver sonograms and are more suitable for ultrasonic liver classification. 2.. Feature Selection for Ultrasonic Liver Texture Classification. This section describes the features employed in our method for classifying the ultrasonic liver images. 2.1 The Selection of the Wavelet Scales In recent years, wavelet transform becomes an active area of research for multiscale signal analysis because it can provide a precise and unifying framework for characterizing a signal at multiple resolution [13]. For wavelet application, the first issue is to select the proper mother wavelet possessing desirable characteristics. The “desirable” characteristics are application dependent. Like Katagishi [14] adopted the fluency wavelet analysis [15] criterion to select the best mother wavelet for detecting characteristic points on signals, such as non-differentiable points and discontinuous ones. In our case, we want to identify the liver status, including the normal liver, hepatitis and cirrhosis, from sonograms. The best mother wavelet should serve to decrease the correlation among the different liver status for the convenience of the successive discrimination task. The mother wavelet candidates in our study include Daubechies [16], least asymmetric [17], Coiflets [18], [19], Battle-Lemarie [20], first derivative of Gaussians, spline-variant [17], and cubic B-spline [21]. Since the first derivative of Gaussians decrease the correlation most significantly, they are adopted as the mother wavelets in this study. The selected mother wavelets are given by. . x2 + y 2 Ψ (x, y) = −x exp − 2 x. and. . 2 x + y2 Ψ (x, y) = −y exp − , 2 y. (1a). (1b). respectively. They also can be viewed as the horizontal and vertical edge enhancement filters. The wavelet transformed value with scale s is equivalent to enhance the liver parenchyma granularity with corresponding size. In addition, we can find that the wavelet profiles are analogous to the basic profiles of the ultrasonic liver images (see Fig. 1) such that they can provide closer correlation between the wavelet transformed results and the liver diseases. Scale selection is another important issue for wavelet application which will affect the subsequent analysis. To determine which scales are to be used, we investigate the wavelet transform and the ultrasonic liver images in the frequency domain. Let ˆ y (ωx , ωy ) denote the Fourier transˆ x (ωx , ωy ) and Ψ Ψ x form of Ψ (x, y) and Ψy (x, y), respectively, ωx2 + ωy2 x ˆ (ωx , ωy ) = j2πωx exp − , (2a) Ψ 2 2 2 ω + ω x y ˆ y (ωx , ωy ) = j2πωy exp − Ψ (2b) 2 as shown in Fig. 2. Similarly, we can derive that the Fourier transform of Ψxs (x, y) and Ψys (x, y), which are scaled 1/s from Ψx (x, y) and Ψy (x, y), are given by ˆ xs (ωx , ωy ) = sΨ ˆ x (sωx , sωy ) Ψ ˆ y (sωx , sωy ). ˆ ys (ωx , ωy ) = sΨ Ψ. (3a) (3b). Fig. 1 Horizontal intensity profiles of ultrasonic liver images. ‘··’ normal liver, ‘- -’ hepatitis, and ‘–’ cirrhosis..
(3) LEE et al.: A NEW APPROACH TO ULTRASONIC LIVER IMAGE CLASSIFICATION. 1303. σωy = =. ∞ 0. √. ˆ y (0, ωy )|2 dωy (ωy − ωy0 )2 |Ψ. π(3π/4 − 2)).. 1/2. (5b). Similarly, it is clear that the rms bandwidth of ˆ xs (ωx , ωy ) and Ψ ˆ ys (ωx , ωy ) are σωx /s and σωy /s, reΨ spectively. Figure 3 shows the sonogram samples of normal liver, hepatitis, and cirrhosis. Figure 4 shows the spectra with dc suppression of the window regions of Fig. 3. Observing these spectra, we find that the spectra in ωy spread to higher frequency even to 0.7π, while the spectra in ωx concentrated in lower frequency about 0–0.3π. To seek the analyzing scales in higher scaleresolution and to cover the above frequency bands, we select scales 0.5, 1.5, 2.5 and 3.5 for Ψy (x, y) while 1, 2 and 3 for Ψx (x, y) for the liver texture analysis in sonograms.. (a). 2.2 Feature Extraction and Selection Inspecting the wavelet transformed liver texture, we find that the intensity variation of different diseases becomes more distinctive after transformation. To measure the intensity variation, four commonly used second-order statistical features, including variance, contrast, covariance and sum variance, are employed. Let ηs be the average value of the transformed image W I(s, x, y) at scale s and (∆x, ∆y) represents the intersample spacing distance vector, where ∆x and ∆y are integers and ∆x = 0 and ∆y = 0. These statistical features are defined as. (b) ˆ x (ωx , ωy ) and Fig. 2 The power spectra of (a) Ψ y ˆ (b) Ψ (ωx , ωy ).. Let (ωx0 , 0) and (0, ωy0 ) be the center of the passing ˆ x (ωx , ωy ) and Ψ ˆ y (ωx , ωy ), respectively, so band of Ψ that ∞ ˆ x (ωx , 0)|2 dωx = 0 (ωx − ωx0 )|Ψ (4a) 0 ∞ ˆ y (0, ωy )|2 dωy = 0. (ωy − ωy0 )|Ψ (4b) 0. By substituting Eq. (2) into Eq. (4), we obtain ωx0 = 2 ωy0 = √ . It is clear that the center of the passing π ˆ x (ωx , ωy ) and Ψ ˆ y (ωx , ωy ) are (ωx0 /s, 0) and bands of Ψ s s (0, ωy0 /s), respectively. Let σωx and σωy be the rms bandwidth around (ωx0 , 0) and (0, ωy0 ), respectively, σωx = =. ∞ 0. √. 2. ˆ x (ωx , 0)|2 dωx (ωx − ωx0 ) |Ψ. π(3π/4 − 2)). 1/2. (5a). variance: (6a) V (s) ≡ E [W I(s, x, y) − ηs ]2 contrast: CON (s, ∆x, ∆y) ≡ E [W I(s, x, y) − W I(s, x + ∆x, y + ∆y)]2 (6b) covariance: COV (s, ∆x, ∆y) ≡ E {[W I(s, x, y) − ηs ] (6c) [W I(s, x + ∆x, y + ∆y) − ηs ]} sum variance: SM V (s, ∆x, ∆y) ≡ E [W I(s, x, y) + W I(s, x + ∆x, y + ∆y)]2 2. −E {[W I(s, x, y) + W I(s, x+∆x, y+∆y)]} (6d) Let I(x, y) denote the intensity of an image at position (x, y). For each transformed image W x I(s, x, y) (= I ∗ Ψxs(x, y)) and W y I(s, x, y) (= I ∗ Ψys (x, y)), both the horizontal and the vertical spacing distance vectors, i.e. (∆x, 0) and (0, ∆y), are used to calculate these statistical features. To determine the upper bounds of the.
(4) IEICE TRANS. INF. & SYST., VOL.E83–D, NO.6 JUNE 2000. 1304. (a). (a). (b). (b). (c) Fig. 4 The Fourier spectra of the windowed regions of Fig. 3 (a)–(c).. (c) Fig. 3 Sampled ultrasonic liver images. The image in (a) is a normal liver. The images in (b) and (c) are hepatitis and cirrhosis, respectively.. distances ∆x and ∆y, we employ the autocorrelation function to measure the correlation level with various distances. The distance, at which the normalized autocorrelation function of the transformed image becomes too small, can serve as an upper bound used for computing these statistical features. Let 0 ≤ x ≤ Lx and 0 ≤ y ≤ Ly denote the considered image region. Let (∆x, ∆y) denote the x-translation and y-translation. The normalized autocorrelation function R(∆x, ∆y) is defined by.
(5) LEE et al.: A NEW APPROACH TO ULTRASONIC LIVER IMAGE CLASSIFICATION. 1305. R(∆x, ∆y) 1 (Lx − |∆x|)(L y − |∆y|) =. ∞. ∞. −∞. −∞. 1 Lx Ly. I(x, y)I(x+∆x, y+∆y)dxdy ∞ ∞ , I 2 (x, y)dxdy −∞. −∞. (7) where |∆x| < Lx and |∆y| < Ly . Let XW x (s) and Y W x (s) denote the distance upper bounds of ∆x and ∆y of W x I(s, x, y). Let XW y (s) and Y W y (s) denote the distance upper bounds of ∆x and ∆y of W y I(s, x, y). By evaluating the normalized autocorrelation functions of the training samples, we select XW x (1) = 3, XW x (2) = 4, XW x (3) = 5, Y W x (1) = 2, Y W x (2) = 2, Y W x (3) = 3, XW y (0.5) = 3, XW y (1.5) = 3, XW y (2.5) = 4, XW y (3.5) = 5, Y W y (0.5) = 2, Y W y (1.5) = 2, Y W y (2.5) = 2, and Y W y (3.5) = 3. The number of extracted features from these wavelet transformed images is 199. To select useful texture features from these extracted features, a multi-dimensional Bayes’ discrimination function combined with the forward sequential search [11] is employed. The method is described as follows: (1) At the first step, the algorithm tests all the M features, and retains the one giving the best classification result as the best single feature F1 . (2) the algorithm continues by combining F1 with each remaining feature. A feature is chosen as the second feature F2 if its combination with F1 optimizes the classification result. (3) additional feature from the remaining feature set is selected to combine with the previously selected best feature set. Repeat this step until the classification capability begins to decrease. The search method has been proven to be effective in saving computation time and reducing possibility of misclassification, since only the most useful features are adopted in liver texture classification. 3.. Liver Tissues Classification. Due to the inherent distributed processing capability, neural networks are frequently employed to classify patterns based on learning from examples. Current methods such as back-propagation use heuristic approaches to discover the underlying class statistics. The heuristic approaches usually involve many small modifications to the system parameters that gradually improve system performance. Besides requiring long computation time for training, the incremental adaptation approach of back-propagation has been shown susceptible to false minima [12]. In addition, retraining is required for the network when new training samples are added. The Probabilistic Neural Network (PNN) [12] can avoid. these problems and possesses other advantages. Its decision surfaces can approach the Bayes optimal. Thus, it is adopted in our algorithm. To accomplish liver classification, the selected texture features of the input pattern X are fed to the PNN. Let fA (X), fB (X), and fC (X) be the outputs of the three summation units of the PNN. To supply quantitative information and indicate the progressive assessment of liver disease to doctors, we modify the decision output as fA (X) fA (X) + fB (X) + fC (X) fB (X) d(XB ) = fA (X) + fB (X) + fC (X) fC (X) d(XC ) = fA (X) + fB (X) + fC (X) d(XA ) =. (8a) (8b) (8c). where d(Xi ) can be viewed as the a posterior probability that pattern X belongs to class i. 4.. Experimental Results. 4.1 Data Acquisition and Feature Selection In this paper, all ultrasonic images are acquired through a Toshiba Sonolayer (SSA250A) with 3.75 MHz transducer at the University Hospital of National ChengKung University. The images are transferred through a VFG frame grabber to a personal computer and are digitized with 512 × 512 pixels and 256 gray-level resolution. The resultant images are then transmitted through the NFS network system to a Sun workstation for further processing. Three sets of ultrasonic liver images including 16 normal samples, 32 hepatitis samples, and 18 cirrhosis samples are tested. All of the samples are from patients receiving liver biopsies. The specimens showed various disease severities from normal to severe cirrhosis. The diagnosis was based on the pathology reports which were reviewed by the researchers. Figure 3 (a) shows one of the normal liver images, while Figs. 3 (b) and (c) show the hepatitis and cirrhosis cases, respectively. For each sampled image, a block of 41 × 41 pixels is selected as the data to be analyzed in the sequel. The block is chosen, if possible, to include solely liver parenchyma without major blood vessels. For each sample block, wavelet transforms are performed. Figure 5 shows the wavelet transformation results, W x I(3, x, y) , of a portion of Fig. 3. The size of the selected portion for wavelet transformation is 256 × 256. Observing these transformed images, we find that the intensity variations are different. Among them, the variation of the cirrhosis image is the most evident. On the contrary, the variation of the normal liver image is not obvious..
(6) IEICE TRANS. INF. & SYST., VOL.E83–D, NO.6 JUNE 2000. 1306. (a). (b). (c). Fig. 5 The wavelet transformation results of the selected portions in Fig. 3 at the third scale along the horizontal direction. Table 1 The classification results of Fig. 3. N = Normal Liver, H = Hepatitis, and C = Cirrhosis.. Table 3. Feature sets used for the classification of ultrasonic liver images.. Table 2 Confusion matrix for the classification of liver sonograms by the proposed method.. For each set of liver images, five training samples are chosen. Then, the forward search scheme is employed to select the useful features from the 199 extracted features. They are variance V (3), contrast CON (3, 3, 0), sum variance SM V (3, 0, 3), and contrast CON (3, 0, 3) all of W x I(3, x, y). These results show that the third scale wavelet transformation along the horizontal direction can enhance the embeddedly inherent difference in granularity of the three considered liver states. And, the difference can be picked up by using some statistical measurements whose intersample spacing distance is three.. 4.2 Classification Results The 66 sampled ultrasonic liver images were classified into the three classes (i.e. normal, hepatitis, and cirrhosis) by using the selected feature vector and the Probabilistic Neural Network. Table 1 shows the classification results of Fig. 3. The confusion matrix using the selected feature vector is shown in Table 2. From this table, we find an encouraging result that there is no misclassification samples between the normal and the cirrhosis sets. This means that our features are powerful for discriminating the cirrhosis from the normal. All the misclassification samples occur at the hepatitis set because its texture structure is between the normal liver and the cirrhosis. The correct classification rate of our method is about 88%. This shows that our selected features are effective for classifying the ultrasonic liver tissues. For comparison purposes, the images are also classified based on other classes of texture features such as the co-occurrence matrices (COOM) [1], the Fourier power spectrum (FPS) [2], the statistical features (SF),.
(7) LEE et al.: A NEW APPROACH TO ULTRASONIC LIVER IMAGE CLASSIFICATION. 1307 Table 4 sets.. The correct classification rate of the tested feature. and the texture spectrum (TS) [3]–[5]. Table 3 shows the selected features of these classes by using the forward sequential search scheme, which can select only the useful features. The actual definitions of these features can be found in the corresponding references. The correct classification rate (CR) using these selected features are shown in Table 4. These results show that our proposed features are superior in ultrasonic liver image classification to these features. In addition, they also demonstrate that our method can gain about eighteen percent advantages over the direct use of the statistical features. This means that the wavelet transform at proper scales can effectively increase the distance of the statistical feature clusters of different liver diseases. 5.. Conclusions. Liver parenchyma with different diseases reveals different granular sizes. In contrast to the conventional approaches to liver classification, we proposed to classify sonogram with different granularity based on a new concept “multiscale” in this paper. Wavelet transforms, a multiscale analysis tool, was exploited to analyze the ultrasonic liver images. From the transformed images, a new set of features reflecting signal variations were derived for classifying three liver states including normal liver, hepatitis and cirrhosis. From the experimental study, we found that all these features were derived from the horizontal wavelet transformation at the third scale W x I(3, x, y). It implies that the third scale can be the representative scale for the classification of the considered liver diseases, and the horizontal wavelet transform can improve the representation of the corresponding features. Experimental results show that our method can achieve about 88% correct classification rate which is superior to other measures by conventional methods, such as the co-occurrence matrices, the Fourier power spectrum, and the texture spectrum. In other words, our method seems more effective in accessing the granularity from sonogram, and thus is capable of classifying the liver tissues more accurately. In addition, it also shows that no misclassification between the normal liver and the cirrhosis sets occurs in the experiments. Therefore, the derived features in. the proposed method are powerful in discriminating the cirrhosis from the normal livers that is the most essential task in liver diagnosis. As a matter of fact, the selected features derived from the wavelet transform are abundant in texture information for liver sonograms, such that only four features are used in our method to achieve so high correct classification rate. The experimental study also provides various evidences of the advantage of the wavelet transform based approach. It not only has better performance than the conventional approaches but also gains eighteen percent increase in correct classification rate over the direct use of the statistical features. Therefore, it suggests that the wavelet transform at proper scales can effectively increase the discrimination capability of the three clusters of different liver states, and can be a potential tool for physicians to quantify liver status in clinical diagnosis. Acknowledgement The authors would like to thank the National Science Council, R.O.C. under Grant No. NSC 89-2213-E-212010 for the support of this work. References [1] R.M. Haralick, K. Shanmugam, and I. Dinstein, “Texture features for image classification,” IEEE Trans. Syst., Man. & Cybern., vol.3, pp.610–621, 1973. [2] G.O. Lendaris and G.L. Stanley, “Diffraction pattern sampling for automatic pattern recognition,” Proc. IEEE, vol.58, pp.198–216, 1970. [3] D.C. He and L. Wang, “Texture features based on texture spectrum,” Pattern Recognition, vol.24, pp.391–399, 1991. [4] L. Wang and D.C. He, “Texture classification using the texture spectrum,” Pattern Recognition, vol.23, no.3, pp.905–910, 1990. [5] D.C. He and L. Wang, “Unsupervised texture classification and segmentation using texture spectrum,” Pattern Recognition, vol.25, no.3, pp.247–255, 1992. [6] L. Carlucci, “A formal system for texture languages,” Pattern Recognition 4, pp.53–72, 1972. [7] S. Tsuji and F. Tomita, “A structural analyzer for a class of textures,” Comput. Graphics Image Process., vol.2, pp.216–231, 1973. [8] J.S. Wesaka, C.R. Dyer, and A. Rosenfeld, “A comparative study of texture measures for terrain classification,” IEEE Trans. Syst., Man. & Cybern., vol.6, pp.269–285, 1976. [9] C.M. Wu and Y.C. Chen, “Multi-threshold dimension vector for texture analysis and its application to liver tissue classification,” Pattern Recognition, vol.26, pp.137–144, 1993. [10] C.M. Wu and Y.C. Chen, “Statistical feature matrix for texture analysis,” CVGIP: Graphical Models and Image Processing, vol.54, no.5, pp.407–419, 1992. [11] D.C. He, Li Wang, and J. Guibert, “Texture discrimination based on an optimal utilization of texture features,” Pattern Recognition, vol.21, pp.141–146, 1988. [12] D.F. Specht, “Probabilistic neural networks,” Neural Networks, vol.3, pp.109–118, 1990. [13] S.G. Mallat, “A theory of multiresolution signal decomposition: The wavelet representation,” IEEE Trans. Pattern Anal. & Mach. Intell., vol.11, pp.674–693, 1989..
(8) IEICE TRANS. INF. & SYST., VOL.E83–D, NO.6 JUNE 2000. 1308. [14] K. Katagishi, T. Urushiyama, K. Toraichi, K. Wada, F. Yoshikawa, and S. Morokami, “On a criterion of selecting the best mother wavelet based on fluency wavelet analysis,” IEEE Conf. pp.485–488, 1997. [15] K. Toraichi, “Fluency theory and wavelet analysis,” The Japan Society of Acoustics, vol.46, no.6, pp.430–436, 1996. [16] I. Daubechies, “Orthonormal bases of compactly supported wavelets,” Comm. on Pure and Appl. Math., vol.41, pp.909–996, 1988. [17] I. Daubechies, Ten Lectures on Wavelets, SIAM, Philadephia, 1992. [18] G. Beylkin, R. Coifman, and V. Rokhlin, “Fast wavelet transforms and numerical algorithms,” Comm. on Pure and Appl. Math., vol.44, pp.141–183, 1991. [19] I. Daubechies, “Orthonormal bases of compactly supported wavelets II. Variations on a theme,” SIAM J. Math. Anal., vol.24, pp.499–519, 1993. [20] G. Battle, “A block spin construction of ondelettes, Part I: Lemarie functions,” Comm. Math. Phys., vol.110, pp.601–615, 1987. [21] M. Unser, A. Aldroubi, and M. Eden, “A family of polynomial spline wavelet transforms,” NCRR Report, National Institute of Health, vol.153/90, 1990.. Jiann-Shu Lee was born in Tainan, Taiwan, Republic of China, on 4 May 1966. He received the BS, MS and Ph.D. degrees in Electrical Engineering from National Cheng Kung University, Tainan, Taiwan, in 1988, 1990 and 1994, respectively. Now he is an assistant Professor at the Department of Computer Science and Information Engineering, Dayeh University, Taiwan. His current research interest is in medical image processing, computer vision, neural network and pattern recognition. Dr. Lee is a member of the Chinese Association of Image Processing and Pattern Recognition, the Institute of Information and Computing Machinery, and the Chinese Association of Biomedical Engineering.. Yung-Nien Sun received the BS degree from National Chiao Tung University, Hsin-Chu, Taiwan, Republic of China, in 1978 and the MS and Ph.D. degrees from University of Pittsburgh, Pittsburgh, Pennsylvania, in 1983 and 1987, respectively. He was Assistant Scientist with the Brookhaven National Laboratory, New York from 1987 to 1989, and he is currently Professor at the Department of Computer Science and Information Engineering, National Cheng Kung University, where he joined in 1989 as an associate professor. He has been working on image processing and computer vision since 1982 and has published more than 120 papers, half of them in referred journals. His current research interests are in medical image analysis, vision, pattern recognition, computer graphics, and Virtual Reality. He is a member of IEEE, Sigma-Xi, the Chinese Association of Image Processing and Pattern Recognition, and the Chinese Association of Biomedical Engineering.. Xi-Zhang Lin received his MD degree from the University of Taiwan and completed internal medicine resident training at Taipei Veteran General Hospital (1983–1986). He completed gastroenterology fellowship at Taiwan University Hospital and then joined National Cheng Kung University (1988). He is the Chief of Division of Gastroenterology at the National Cheng Kung University Hospital and an Associate Professor at Medical College. His main research interests include medical imaging on hepatology and Gastroenterology, especially in the quantitation of the information from medical imagings, such as imaging analysis of liver echotexture, three-dimensional reconstruction of the organs, and image-guided therapy..
(9)
數據
相關文件
In this paper, we extended the entropy-like proximal algo- rithm proposed by Eggermont [12] for convex programming subject to nonnegative constraints and proposed a class of
In this paper, we build a new class of neural networks based on the smoothing method for NCP introduced by Haddou and Maheux [18] using some family F of smoothing functions.
Chen, The semismooth-related properties of a merit function and a descent method for the nonlinear complementarity problem, Journal of Global Optimization, vol.. Soares, A new
Section 3 is devoted to developing proximal point method to solve the monotone second-order cone complementarity problem with a practical approximation criterion based on a new
In summary, the main contribution of this paper is to propose a new family of smoothing functions and correct a flaw in an algorithm studied in [13], which is used to guarantee
The purpose of this talk is to analyze new hybrid proximal point algorithms and solve the constrained minimization problem involving a convex functional in a uni- formly convex
Huang, A nonmonotone smoothing-type algorithm for solv- ing a system of equalities and inequalities, Journal of Computational and Applied Mathematics, vol. Hao, A new
Bell’s theorem demonstrates a quantitative incompatibility between the local realist world view (à la Einstein) –which is constrained by Bell’s inequalities, and