Académique Documents
Professionnel Documents
Culture Documents
I. INTRODUCTION
(4)
(5)
Fig. 3. The division of spectrum after two iterations of the traditional dyadic
where . In the wavelet literature [16], wavelet decomposition.
[17], many different sets of coefficients can be found,
corresponding to wavelet bases with different properties. In
the case of the discrete wavelet transform (DWT), coefficients
play an important role, since they can be used for the
DWT computation, instead of the explicit forms for and
. It is shown [18] that, starting from the original signal
, discrete signals [the approximation
of at resolution ] and (information content lost
between higher resolution and lower resolution ) can Fig. 4. Filter bank performing the quincunx wavelet transform.
be computed as
are
(7a)
(7b)
(6)
Since this transform is performed with the two-
channel filter bank shown in Fig. 4. The Fourier expression
where and . This decomposition for the output of channel is
can be understood as passing of signal through a pair
of low-pass and high-pass filters and , followed by
the subsampling with a factor two.
image is decomposed with the multiresolution scale factor . and to map coefficients of the selected 1-D filter into a 2-D
This is very nice property for description of small textured filter defined on the quincunx lattice. The transform obtained
images, since the analysis is twice as fine as the separable ensures that all properties of 1-D filters are also satisfied in the
multiresolution decomposition. The spectral decompositions 2-D case. Derivation of the transform is given in the Appendix.
shown in Figs. 3 and 5 suggest several advantages of the
quincunx transform for tissue characterization from B-scan VI. TISSUE CLASSIFICATION
images. First, the separable sampling provides only rectangular
divisions of spectrum, with increased sensitivity to horizontal A. Data Acquisition
and vertical edges. This could be important for the analysis
of directional textures, but it yields a rotationally sensitive The ultrasound images used in this research were obtained
description, which is not desirable in this application. Due to on Toshiba SSA-100 equipment, with 3.5-MHz transducer
the shape of the low-pass and high-pass filters, it should be frequency. A Series 151 Image processor, from Imaging
expected that the quincunx decomposition has lower orienta- Technology Inc., on IBM-PC, was used to capture images with
tion sensitivity than separable decomposition. Still, it is not 512 512 pixels and 256 gray-level resolution. Three sets of
completely rotationally insensitive—following the rotation of images have been taken: normal (37 images from ten subjects),
the QT Voronoi cell around the origin, rotational sensitivity steatosis (65 images from ten patients), and cirrhosis (20
increases up to 45 (where it reaches maximum), and then images from ten patients). Since this study addresses the issue
decreases again, reaching the complete invariance for 90 . of diagnosing early cirrhosis and steatosis, we have constrained
Second, the energy of natural textures is mainly concentrated the selection of images to those from patients in early stage
in the mid-frequencies, with the insignificant energy along of the disease. From each image, two blocks of 64 64 pixels
diagonals. Therefore, the quincunx low-pass filter will preserve (approximately 2 cm 2 cm in actual dimensions) have been
more of the original signal energy, and its implementation in selected. Blocks were chosen to include only liver tissue,
the iterated filter bank could provide more reliable description without blood vessels, acoustic shadowing, or any type of
of texture. Finally, the diamond shape of the low-pass filter distortion. In that way, the whole data set contained 244 tissue
in the quincunx case, plays the crucial role for the extraction samples for training and classification. The training and test
of texture features in the presence of noise, since it cuts off sets (each with 122 samples) were composed out of all blocks
diagonal high frequencies, where the most significant portion from independent images.
of noise is contained. Thus, when working with noisy samples In patients with steatosis and cirrhosis, the final diagno-
(as in our case) the spectral decomposition performed on the sis was confirmed by liver biopsy and histology, since this
quincunx lattice represents a better solution than traditional presents the gold standard for diagnosing diffuse diseases. The
approach based on the separable sampling. needle diameter was 1.6 mm, and the obtained tissue cylinders
were at least 2 cm.
The estimation of a texture quality was performed with the
V. DESIGN OF 2-D DIAMOND-SHAPED FILTERS four-level quincunx decomposition, yielding feature vectors
When constructing a filter bank performing the DWT for with maximal length of five. The size of the smallest subimage
the texture characterization, a number of design requirements in the quincunx pyramid is 16 16 pixels and further decom-
have to be fulfilled. First, since images are mostly smooth, the position would yield unreliable estimates of texture quality.
analysis should be performed with a smooth mother wavelet. Fig. 6 shows the first three levels of the quincunx pyramid,
On the other hand, to achieve fast computation, the filters have for one representative sample from each tissue class. The
to be short, affecting the smoothness of the associated wavelet. decomposition was performed by using the three biorthogonal
In more general applications, such as texture synthesis, it families that provide short filters with the best tradeoff between
would be nice to construct a perfect reconstruction filter bank, the filter lengths and the regularity of corresponding wavelet.
MOJSILOVIC et al.: CHARACTERIZATION OF VISUALLY SIMILAR DIFFUSE DISEASES 545
C. Classification
Assuming that feature vectors for normal, cirrhosis, and
steatosis tissue have approximately normal distribution (with
mean vector and covariance matrix ), the minimum error
(a)
discriminant function between the feature vector and class
is the Bayes distance
(13)
(b)
For each texture class, the mean and the covariance were
estimated using 122 texture samples with the leave-one-out
method [24]. The remaining samples were then classified.
(c)
Fig. 6. Quincunx decomposition of: (a) cirrhosis tissue, (b) normal tissue,
VII. RESULTS
and (c) steatosis tissue. The classification results obtained using all features of
quincunx decomposition are presented in Table III. The overall
A brief description and 1-D filter coefficients for each family classification accuracy is 90%. Specificity of the method was
are given in Table I. Table II presents coefficients of the 92%. Sensitivity in the detection of cirrhosis and steatosis
corresponding diamond-shaped filter. To illustrate the 1-D to were: 92 and 97%, respectively. These results correspond to
2-D mapping, the frequency responses of the second family decomposition with the second pair of filters (see Table II,
and its diamond-shaped counterparts are shown in Fig. 7. set 2). Similar results were obtained with the other two
sets, indicating that the texture classification scheme in this
application is almost insensitive to the selection of the wavelet
B. Texture Feature Extraction base.
Since the filter bank performing the QT represents one spe- As already mentioned in the data-acquisition section, the
cial case of the local linear transform approach for the texture training and classification sets were formed from independent
characterization, iterations of the quincunx decomposition images. Therefore, it is to be expected that in the classification
can be seen as a -channel filter bank, whose outputs phase both blocks from one image should be diagnosed
serve for the estimation of texture quality identically. This is true for normal and steatosis classes, but
in the corresponding frequency subband. The texture is then in the cirrhosis case, one block from an image was diagnosed
characterized by the set of first-order probability density differently from the other. More detailed examination showed
functions estimated at the output of each channel. More that this one, as well as many other misclassified samples was
compact representation can be achieved using the channel affected by noise. Some of them had inhomogeneous structure
variances var . The statistical justification for this approach due to small blood vessels present in that part of liver. Hence,
can be found in [20]. Another, psychophysical justification was it is obvious that the location of the region of interest (ROI)
offered by Pratt et al. [21], who showed that natural textures within an image has dominant effect on the classification. This
are visually indistinguishable if they possess the same first issue is addressed in more detail in Section VIII.
and second-order statistics. Also, the reliability of channel
variances has been proved in numerous wavelet-based texture A. Comparison Between Separable and
characterization algorithms [11], [12], [22], [23]. Since the Nonseparable Transform
high-pass conditions and imply
To compare the performance of the nonseparable quincunx
that , for , channel variances are
decomposition with the performance of the separable wavelet
represented through energies calculated at
transform, we have also performed two-level ST with the
the output of each channel with
same filter family (Table I, set 2), on the same data set. This
decomposition results in maximum of seven features and the
(12) same size of the smallest subimage. Using the same method
for training and classification and using the complete feature
set resulted in an 88% of overall classification accuracy. By
decreasing the number of features from five to three (for
where represents the image field, # denotes the number QT), and from seven to three (for ST), the classification
of pixels in , and stands for wavelet rate remained constant in the quincunx case, whereas in the
coefficient subimages obtained at the th decomposition level. ST case the overall classification rate decreased to 82%. By
Finally, due to the low-pass condition , to obtain inspection of misclassified samples, we have found that almost
the variance in the last low-pass channel, the mean of all samples that were misclassified with the QT, were also
the image should be removed. misclassified using the ST.
546 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 17, NO. 4, AUGUST 1998
TABLE I
FILTER COEFFICIENTS FOR THE 1-D BIORTHOGONAL WAVELET BASIS USED FOR THE DESIGN OF 2-D
DIAMOND-SHAPED FILTERS. THE FIRST FAMILY BELONGS TO THE SO-CALLED “SPLINE FILTERS” OR
“BINOMIAL FILTERS” IN [15], BECAUSE THE UNDERLYING SCALING FUNCTION IS A B-SPLINE. THE SECOND
FAMILY IS ALSO A “SPLINE VARIANT” BUT WITH LESS DISSIMILAR LENGTHS (AS OPPOSED TO BINOMIAL
FILTERS). THE EXAMPLE WE HAVE USED HAS THE SHORTEST FILTERS WITHIN
THIS FAMILY. THE THIRD SET OF COEFFICIENTS BELONGS TO FILTERS CLOSE TO
ORTHONORMAL FILTERS WHERE SCALING FUNCTION AND WAVELETS ARE VERY SIMILAR
TABLE II
FILTER COEFFICIENTS FOR THE 2-D BIORTHOGONAL DIAMOND-SHAPED FILTERS
TABLE III
CLASSIFICATION RESULTS OBTAINED WITH THE QUINCUNX DECOMPOSITION
(a)
B. Sensitivity to Rotation
We were also interested in testing the ability of nonseparable
decomposition to classify textures with respect to rotation.
In the second experiment, without changing already esti-
mated class prototypes, we have classified texture samples
rotated at 5 , 23 , 45 , and 90 . Samples for the experiment
were obtained in the following manner. First, the whole
image was rotated for the determined degree, and then two
tissue samples were taken from the same positions as in
(b) the previous experiment. The rotation was performed using
Fig. 7. Magnitudes of the frequency response for: (a) 1-D low-pass and a bicubic interpolation. The QT yielded overall accuracy
high-pass prototype filters and (b) their 2-D counterparts. The filters corre- of 88%, 88%, 80%, and 90% (for 5 , 23 , 45 , and 90 ,
spond to biorthogonal spline families with low-pass and high-pass filters of
similar length. The coefficients of 1-D prototypes correspond to the second respectively), whereas ST gave 82%, 81%, 56%, and 81%.
family in Table I. These results follow from the analysis in Section IV. It is
MOJSILOVIC et al.: CHARACTERIZATION OF VISUALLY SIMILAR DIFFUSE DISEASES 547
obvious that the quincunx decomposition performs better wavelet transforms are very promising. Based on the ex-
than the separable decomposition, but the complete rotation perimental results, we conclude that the quincunx transform
invariance is not achieved since the classification accuracy is is more appropriate for characterization of noisy data, and
not maintained at different angles. One way to achieve the for practical applications requiring description with lower
complete invariance is to use any circularly shaped spectral rotational sensitivity. The results obtained in the study, clearly,
decomposition, such as hexagonal decomposition, steereable illustrate the utility of sophisticated image processing algo-
pyramid, or Gabor transform [25], [26]. Each of these trans- rithms in the ultrasound diagnostics of liver. However, for
forms has very fine orientation selectivity and produces a the general application of the method, many issues still have
number of features that correspond to a certain orientation. to be developed. It is obvious from our results, as well
Hence, for the selected transform, by averaging all features as from the previous work [10], [27], that the location of
from the same resolution level, we will produce one rota- the ROI within an image has the dominant effect on the
tionally insensitive feature. Unfortunately, by achieving the classification. This is due to the fact that ultrasound pattern
rotational insensitivity, we are loosing valuable features in exhibits different behavior along the path of the acoustic pulse,
terms of general description, and the overall accuracy of resulting in the prominently different distribution of horizontal
the method can decrease. Another solution would be the edges along axis of an image. Also, the lateral speckle pattern
construction of tissue prototypes for different orientations. is strongly depth dependent [27]. Finally, echocardiogram is
For example, instead of one cirrhosis class we would have formed according to the sector, while images are sampled
classes, for different degrees of orientation. In that case, along the rectangular grid. During the selection of tissue
instead of classification into three groups, we would perform samples we have done our best to choose the data set as
classification into groups. Although promising, for reliable homogenous as possible. This was done by taking the samples
estimation of class prototypes, this approach requires a huge from an isolated region in the center of an image without
amount of data and even more sophisticated feature-extraction great changes in the cursor position, but our selection was
techniques. severely constrained by the presence of major blood vessels
and acoustic shadowing. It should be also mentioned that the
C. Comparison with Other Texture Description Methods images were recorded by different clinicians who were free
to adjust ultrasound settings (within the acceptable range) in
To compare the performance of different feature-extraction
order to visually optimize images. We feel that this degree
techniques in diagnosing early diffuse diseases, we have
of freedom is required, so the whole method could follow
compared the wavelet-based approach with: 1) gray-level
natural processing of each individual physician, but it is to be
concurrence (GLC) matrices, as defined in [7], and [10]; 2)
expected that a more rigorous selection of tissue samples could
fractal texture measures [7]; and 3) Fourier measures [7].
increase classification accuracy. At this point, the advantages
In this experiment, using the same data set, the following
of the wavelet approach over other methods become obvious.
measures have been calculated.
First, since the mean values of all DWT channels are zero, the
1) Contrast, angular second moment, and correlation for the changes in the gain setting could not affect the results. Also,
GLC method [7], [10]. The GLC matrices were formed it has been shown that the wavelet transform is powerful ana-
for all combination of displacement and lytical tool for assessment of the singularities [16], [17]. Since
and angles and . visual differences between the early diffuse diseases are hardly
2) Ring and wedge energies, derived from distinguishable, due to minor changes in tissue structure, the
the Fourier power spectrum [6], for all combinations wavelet transform is appropriate for “zooming” into those dif-
and ferences. Furthermore, the ultrasound noise is a high-frequency
. phenomenon, mostly affecting a single image point. Hence, by
3) Multithreshold fractal features proposed for liver tissue eliminating from the classification high-frequency channels at
classification as defined in [7]. the highest resolution level, the impact of ultrasound noise
The highest overall classification accuracy was: 87% for the can be significantly reduced. Finally, the proposed feature
GLC method, 82% for Fourier measures, and 69% for fractal method is independent of the classification technique. The
measures, indicating the superiority of the wavelet approach. wavelet-based texture measures can be used with minimum-
distance-type classifiers, Bayesian-type classifiers, or as inputs
to a neural network. Since previous investigations showed that
VIII. DISCUSSION AND CONCLUSION the classification technique has significant influence to the final
The goal of this research was the detection of an optimal diagnosis [8], [10], further research should be performed in
feature-extraction technique for description of different ultra- that direction.
sound textures. Therefore, we have investigated the utility
of wavelet decompositions as feature-extraction methods in
discrimination among diffuse liver diseases. We have applied APPENDIX
a nonseparable quincunx transform with a multiresolution DERIVATION OF THE MCCLELLAN TRANSFORM
FOR DESIGN OF DIAMOND-SHAPED FILTERS
scale factor , traditional approach based on the separable
wavelet transform, and compared them with the previously The mapping of a 1-D zero-phase symmetric filter
used approaches. The classification results obtained with both into a 2-D filter starts with the Fourier
548 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 17, NO. 4, AUGUST 1998
(A2)
(A10)
where , , and , and
where corresponds to the Voronoi cell of the quincunx
is the length of the filter. The next step is to write
lattice. In this particular case we minimize under
as a polynomial of . This can be achieved
the constraints (A7) and (A9). Consequently, the minimization
by using the Chebyshev polynomials, since can be
of the error function is the linear problem. Hence, we obtain
written as , where is the th-order Chebyshev
the simplest, first-order transform for a design of diamond-
polynomial. Hence
shape filters
(A11)
(A3)
REFERENCES
The idea of the McClellan transform is to replace by a [1] P. N. Wells, Biomedical Ultrasonics. New York: Academic, 1977.
zero-phase 2-D filter . This will result in an overall [2] R. Kuc and M. Schwartz, “Estimating the acoustic attenuation coefficient
zero-phase 2-D filter slope for liver from reflected ultrasound signals,” IEEE Trans. Sonics
Ultrason., vol. SU-26, pp. 353–362, Sept. 1979.
[3] R. Kuc, “Processing of diagnostic ultrasound signals,” IEEE ASSP Mag.,
pp. 19–26, Jan. 1984.
(A4) [4] R. Momenan, M. F. Insana, R. Wagner, B. S. Garra, and M. H.
Loew, “Application of cluster analysis and unsupervised learning to
multivariate tissue characterization,” J. Clin. Eng., vol. 13, pp. 455–461,
Typically is chosen as 1988.
[5] 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
(A5) the Mahalanobis distance,” IEEE Trans. Med. Imag., vol. 13, pp. 37–47,
Mar. 1994.
[6] D. Nicholas, D. K. Nassiri, P. Garbutt, and C. R. Hill, “Tissue char-
acterization from ultrasound B-scan data,” Ultrasound Med., Biol., vol.
Therefore, the transform maps the frequency re- 12, no. 2, pp. 135–143, Feb. 1986.
sponse of a 1-D filter to the contours in plane. The [7] C. Wu, Y. Chen, and K. Hsieh, “Texture features for classification of
ultrasonic liver images,” IEEE Trans. Med. Imag., vol. 11, pp. 141–152,
shape of the obtained contours is determined only by the June 1992.
transformation parameters , which give the structure of [8] M. S. Klein Gebbinck, J. T. M. Verhoeven, J. M. Thijssen, and T.
a function . In this section we will consider the problem E. Schouten, “Application of neural networks for the classification of
diffuse liver disease by quantitative echography,” Ultrason. Imag., vol.
of choosing the transformation parameters so that the contours 15, no. 3, pp. 205–217, July 1993.
produced by the transformation, for various const, have a [9] H. Paik and M. D. Fox, “Fast Hartley transforms for image processing,”
diamond shape in plane. In that case, the coefficients IEEE Trans. Med. Imag., vol. 7, pp. 149–153, June 1988.
[10] Y. M. Kadah, A. A. Farag, J. M. Zurada, A. M. Badawi, and A. M.
for the low-order transformation can be determined as follows. Youssef, “Classification algorithms for quantitative tissue characteriza-
First, we have to impose the structure of the transformation. tion of diffuse liver disease from ultrasound images,” IEEE Trans. Med.
The simplest form for , in accordance with the Imag., vol. 15, no. 4, pp. 466–477, Aug. 1996.
[11] A. Mojsilović, M. Popović, and D. Šević, “Classification of the ultra-
lattice points of the quincunx sampling scheme, is 2
sound liver images with the 2 1-D wavelet transform,” in Proc. 1996
IEEE Int. Conf. Image Processing, Sept. 1996, vol. 1, pp. 367–370.
(A6) [12] A. Mojsilović, M. Popović, A. Nešković, and A. Popović, “Wavelet
image extension for analysis and classification of infarcted myocardial
tissue,” IEEE Trans. Biomed. Eng., vol. 44, pp. 856–866, Sept. 1997.
Since our objective is to design a 2-D low-pass filter based on [13] E. Dubois, “The sampling and reconstruction of time varying imagery
the 1-D low-pass prototype, we must require that 1-D origin with application in video systems,” Proc. IEEE, vol. 73, pp. 502–522,
maps to 2-D origin. This gives the first constraint to the filter Apr. 1985.
[14] J. Kovacevic and M. Vetterli, “Nonseparable multidimensional perfect
reconstruction filter banks and wavelet bases for Rn ,” IEEE Trans.
(A7) Inform. Theory, vol. 38, pp. 533–555, Mar. 1992.
[15] E. Viscito and J. P. Allebach, “The analysis and design of multidi-
In the context of wavelet-based filter banks, we would like to mensional FIR perfect reconstruction filter banks for arbitrary sampling
transform a 1-D filter into a 2-D one, such that zeros at aliasing lattices,” IEEE Trans. Circuits Syst., vol. 38, pp. 29–41, Jan. 1991.
[16] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: SIAM,
frequencies , are preserved 1992.
[17] M. Vetterli and J. Kovacevic, Wavelets and Subband Coding. Engle-
(A8) wood Cliffs, NJ: Prentice-Hall, 1995.
MOJSILOVIC et al.: CHARACTERIZATION OF VISUALLY SIMILAR DIFFUSE DISEASES 549
[18] S. Mallat, “A theory for multiresolution signal decomposition: The [23] T. Chang and C. Kuo, “Texture analysis and classification with tree-
Wavelet representation,” IEEE Trans. Pattern Anal. Machine Intell., vol. structured wavelet transform,” IEEE Trans. Image Processing, vol. 2,
11, pp. 674–693, July 1989. pp. 429–441, Oct. 1993.
[19] R. M. Mersereau, W. F. G. Mecklenbräuker, and T. F. Quatieri, “Mc- [24] R. O. Duda and P. E. Hart, Pattern Classification and Scene Analysis.
Clellan transformation for two-dimensional digital filtering: I-design,” New York: Wiley, 1973.
IEEE Trans. Circuits Syst., vol. 23, pp. 405–413, July 1976. [25] R. Mersereau, “The processing of hexagonally sampled two-dimensional
[20] M. Unser, “Local linear transforms for texture measurements,” Signal signals,” Proc. IEEE, vol. 67, pp. 930–949, June 1979.
Processing, vol. 11, pp. 61–79, July 1986. [26] E. P. Simoncelli, W. T. Freeman, E. H. Adelson, and D. J. Heeger,
[21] W. Pratt, O. Faugeras, and A. Gagalowitz, “Visual discrimination of “Shiftable multiscale transforms,” IEEE Trans. Inform. Theory, vol. 38,
stochastic texture fields,” IEEE Trans. Syst., Man, Cybern., vol. SMC-8, Mar. 1992.
pp. 796–804, 1978. [27] B. J. Oosterveld, J. M. Thijssen, and W. A. Verhoef, “Texture of B-mode
[22] M. Unser, “Texture classification and segmentation using wavelet echograms: 3-D simulations and experiments of the effects of diffraction
frames,” IEEE Trans. Image Processing, vol. 4, pp. 1549–1560, Nov. and scatterer density,” Ultrason. Imag., vol. 7, no. 2, pp. 142–160, Apr.
1995. 1985.