- Email: [email protected]
Noninvasive diagnosis of cardiac rejection through echocardiographic tissue characterization
Noninvasive Diagnosis of Cardiac Rejection Through Echocardiographic Tissue Characterization Evelin Lieback, MD, Rudolph Meyer, MD, Michael Nawrocki, Jochen Bellach, and Roland Hetzer, MD German Heart Institute, and Medical Department, Humboldt University, Berlin, Germany
Ultrasonic tissue characterization is based on the assumption that microscopic tissue structures are identifiable by their acoustic properties. Our study group consisted of 23 cardiac recipients. Two-dimensional images were obtained within 2 hours of endomyocardial biopsy. The end-diastolic echo frames were digitized into the matrix of an image-processing system. A region of interest was placed into the anteroseptal segment of the left ventricle. The texture within the region of interest was analyzed using four major groups of texture analysis (first-qrder histogram, co-occurrence matrix, run-length statistic, and power spectrum). A total of 408 echocardiographic examinations were compared with histologic
findings. The 117 initially calculated texture parameters were reduced incrementally using a series of discriminant analyses. A set of three texture parameters (inverse difference moment undirected, run-length nonuniformity vertical, and sector sum) was able to describe changed echocardiographic texture when rejection occurred. Using these three parameters, echocardiographic sensitivity was 89.0% and specificity was 83.6% for moderate rejection. We conclude that cardiac rejection is associated with echocardiographic texture alterations and that serial echocardiographic texture analysis can reliably identify rejection. (Ann Thorac Surg 1994;57:1168&70)
T
Postoperative Management
o date, echocardiographic diagnosis of cardiac rejection has focused mainly on the analysis of functional, particularly diastolic, parameters of the heart [l]. The purpose of this study was to evaluate the ability of texture analysis of echocardiograms to detect morphologic changes in the myocardium that were caused by rejection after cardiac transplantation. This research is based on the assumption that abnormalities in microscopic tissue structure are reflected in its acoustic properties. Current approaches to ultrasonic tissue characterization include measurement of acoustic parameters such as attenuation coefficients [2] and integrated backscatter [3] and digital image analysis of two-dimensional echocardiograms [4]. Ultrasonic tissue characterization research has shown that these techniques can identify acute ischemic [5], reperfused [6], infarcted [7], and inflamed [8] myocardium. This study examined whether rejection is associated with an abnormal pattern of echocardiographic texture and whether the evolution of the disease can be recognized by means of quantitative texture analysis.
Material and Methods The study group consisted of 23 patients (3 women and 20 men) who had had orthotopic cardiac transplantation. Mean age was 37.2 & 9.0 years (range, 19 to 48 years). Follow-up ranged from 3 to 16 months. Accepted for publication Aug 12, 1993. Address reprint requests to Dr Lieback, German Heart Institute, Augustenburger Platz 1, 13353 Berlin, Germany.
0
1994 by The Society of Thoracic Surgeons
All patients received standard triple-drug immunosuppression therapy with cyclosporine, azathioprine, and prednisone. Cytolytic therapy with antithymocyte globulin was given on postoperative days 1 to 3. Cardiac rejection was monitored using endomyocardial biopsy performed weekly during the first postoperative month and biweekly during the second and third months. Biopsy intervals became longer later in the postoperative course. Biopsy specimens were graded according to the classification of Billingham [9]. Moderate and severe episodes of rejection were treated with pulsed methylprednisolone (0.5 g/day for 3 days). In the case of persistent rejection, rabbit antithymocyte globulin, or OKT3 if necessary, was administered over another 3 days.
Standardized Echocardiographic Examination and linage Digitization Echocardiographic examinations were performed within 2 hours of endomyocardial biopsy without knowledge of the histologic result. The examinations were carefully standardized. Two-dimensional echocardiograms were produced using an electronic 90-degree sector-scanning system operating at 3.75 MHz. Short-axis images were obtained at the midventricular papillary muscle Level. Special care was taken that the angle of the interrogating ultrasound wave in the anteroseptal segment was approximately 90 degrees. Amplification adjustment was standardized for each patient, ie, the first amplification value used on each patient was maintained throughout the study. In general, amplifications between 72 and 76 dB were used. The other variables, such as dynamic range, enhancement, reject, and gamma correction, were kept 0003-4975/94/$7.00
LIEBACK ET AL ECHOCARDIOGRAPHICTISSUE CHARACTERIZATION
Ann Thorac Surg 1994;571164-70
ZEE
I
1165
occurrence matrix, run-length analysis, and power spectrum (Appendix 1).
Intensity
Data Analysis and Classification
Pixel
A
25
Intensity
25
Analysis of ultrasonic texture data is performed to determine which combination of texture variables is necessary and adequate for distinguishing between various histologic conditions of the myocardium (rejection versus no rejection) after cardiac transplantation. The 117 initially calculated texture parameters were reduced to an adequate number using a series of discriminant analyses. This parameter selection process was done to obtain a number of parameters that was relatively low and in which erroneous classification was relatively rare (see Appendix 1). After calculating the classification principle ("optimal" texture parameter set), the computer classified the data from the texture analysis of the 408 echocardiograms obtained during the study. Sensitivity, specificity, and accuracy of texture analysis in comparison with the endomyocardial biopsy results were calculated according to the following formulas: Sensitivity = TP/(TP + FN) Specificity = TN/(FP + TN)
Pixel
25
B Fig 1. Example of three-dimensional pixel maps of gray-level distribution in the region of interest of a representative patient in the phase without rejection and at the time of rejection. Note (A) the relatively homogeneous gray-level values in the phase without rejection and ( B ) the high differences in gray-level values at the time of rejection.
constant. Depth control was standardized at 15 cm. During the study, a total of 408 echocardiographic examinations were conducted. The electrocardiogram-triggered end-diastolic images were digitized into the imageprocessing system.
Endomyocardial Biopsy A median of five tissue samples were taken by endomyocardial biopsy. Biopsy specimens were considered positive when two independent examiners arrived at the same results indicating either moderate or severe rejection.
Texture Analysis The pattern of gray-level distribution in an image is called texture. Echocardiographic texture is defined as regional distribution of echo amplitudes, which are projected on the screen as gray levels. Regions of interest were chosen by using an interactive computer program. These areas, having a standard size of 25 x 25 pixels, were placed into the anteroseptal segment of the left ventricle. The texture within the region of interest was analyzed using measurements related to intensity and distribution of echo amplitudes (gray levels). In the present study, four major types of texture analysis were used: first-order histogram, co-
Accuracy = (TP + TN)/(TP + FN + FP + TN), where TP = true-positive, TN = true-negative, FP false-positive, and FN = false-negative.
=
Results Echocardiographic texture of the myocardium changed significantly when rejection began to develop. Figure 1 shows the gray-level distribution in the region of interest of a representative patient before rejection and during rejection. For quantification of gray-level distribution, 117 texture parameters were calculated. The first 131 echocardiographicexaminationsand the corresponding endomyocardial biopsy specimens were used as the learning group. Analysis of the myocardial biopsy samples produced the following results: no rejection, 54; rejection, 39; and fibrosis, 38. The 117 texture parameters were incrementally reduced by means of upward and downward processing of discriminant analysis. The following parameters were found sufficient to effectively distinguish between rejection, fibrosis, and no rejection: sector sum 0 degrees to 30 degrees; run-length nonuniformity vertical; inverse differ-
+ ++ + ,
4.38
L++++%+ M 1
B I
7.77
OD
M r
3
0:
* 11.2
* **
o&X* ,
M
*$*
8
2 14.6
Fig 2. Distribution of the first nonelementary discriminant function of the learning class showing the ability to distinguish between "rejec= l mean value; + = rejection; * = no tion'' and "no rejection." & rejection; 0 = fibrosis.)
1166
LIEBACK ET AL ECHOCARDIOGRAPHIC TISSUE CHARACTERIZATION
ence moment (dx = 2 undirected). The distribution of the observations concerning the first nonelementary discriminant functions and their classification in no rejection, rejection, and fibrosis is shown in Figure 2. The risk of classification according to Bayes amounted to 0.108425. After the classification principle (texture parameter set) had been found by discriminant analysis, the parameter set was used to classify all 408 data sets of texture analysis. These data sets were compared with their corresponding histologic examination data sets for variances in sensitivity, specificity, and accuracy. Results of no rejection and fibrosis were added to indicate "biopsy negative." The results are presented in Table 1. The sensitivity of the computer classification was 89% and the specificity was 83.6%. Single-case analyses demonstrated that texture parameters can indicate rejection early. Therefore, it is very important to note the changes in echocardiographic texture after cardiac transplantation. The postoperative course of a selected patient is presented in Figure 3. In this patient, the first three biopsy controls were negative, and echocardiographic texture at this point was normal. On the 31st postoperative day, echocardiographicimage texture began changing, and among the changes was a decrease in run-length nonuniformity. The biopsy specimen did not yet indicate rejection. However, a subsequent biopsy specimen 7 days later showed signs of moderate rejection. The run-length uniformity continued to be low. Immunosuppression therapy was applied. After 2 weeks, further examination revealed normal histologic conditions. Echocardiographic texture also returned to normal. During the later postoperative course, texture parameter changes coincided with biopsy findings of rejection.
Comment In previous attempts to diagnose cardiac rejection echocardiographically, quantitative analysis of the effects of rejection-induced changes in organic function was aimed at factors such as increased ventricular wall thickening and left ventricular muscle volume. These changes indicated early manifestation of myocardial edema in connection with rejection before the introduction of cyclosporine. When therapy was combined with cyclosporine, myocardial edema became less pronounced. Because patients receiving cyclosporine often had development of
Table 1. Computer Classification of "Rejection" and "No Reiection'" Computer Diagnosis Biopsy Diagnosis Rejection No rejection
Rejection
No Rejection
75 32
9 272
Using a set of three texture parameters (inverse difference moment, run-length nonuniformity, and sector sum), the echocardiographic sensitivity for rejection was 89.0%, the specificity was 83.6%, and the accuracy was 85.0%.
a
Ann Thorac Surg
1994;57116470
Run-length nonunilorrnlly
80 75 70 65 60 55
50 45 40 35 30 25 20 15 10
5
n 7
BIOPSY +o
TREATMENT
31
38
52
3 tk t 0
no
14
21
moderate
***
A
3 x 0.5 aMP
08
A
no
88
108 115
A
0
no A moderatgo
*.*
3 x 0.5 aMP
135
A
no
102 170 days ~ o . 1 HTx
0' **.
3 I 0.5 a M P
Fig 3. Postoperative course of a representative patient. The first three biopsy controls were negative, and echocardiographic texture was; normal. On the 31st day, the run-length nonuniformity decreased, but endomyocardial biopsy sample did not indicate rejection. The biopsy control 7 days later show signs of moderate rejection. Run-length nonuniformity continued to be low. After immunosuppression therapy, an examination revealed normal histologic conditions. Echocardiogrcrphic texture also returned to normal. (HTx = cardiac transplantation; MP = methylprednisolone.)
arterial hypertension with early left ventricular hypertrophy, wall thickening or muscle volume, which is deduced from it, began to lose diagnostic value. Other researchers have examined to what extent acute rejection is able to cause changes in systolic and diastolic cardiac function. During the rejection process, few or no changes in systolic function were found. However, left ventricular diastolic dynamics are affected by alterations in isovolumetric relaxation, early diastolic filling, anld left ventricular distensibility [lo, 111. Impaired diastolic cardiac function can be explained as a result of increased stiffness through interstitial cellular infiltration and myocardial edema, thus leading to a reduction in compliance of the left ventricle. Other researchers occasionally encountered false-negative results, even in cases of severe rejection, thus making assessment of diastolic ventricular function alone inadequate for reliable diagnosis. In this study, acoustic properties of cardiac allografts were assessed by measuring the texture parameters of echocardiograms. Although the relationship between echocardiographic image texture and anatomic composition of tissue is complex and not yet completely uinderstood, one can assume that the processes connected with rejection, such as cellular infiltration, myocytolysis, linterstitial edema, and changed perfusion, influence the acoustic properties of the myocardium and, with this, the echocardiographic image texture. The distance between the scatters is enlarged by interstitial edema, ie. the number of scatters per volume unit diminishes. The increased liquid content of the tissue, on the other hand, influences attenuation qualities, ie, increased water content decreases attenuation. The density and elasticity of the myocardium are changed by cellular infiltration, a
Ann Thorac Surg 1994;571164-70
condition usually causing increased reflection. At the same time, the size and shape of scatters change. Changes in fiber orientation and rearrangement of dominant myofibrils occur. The edges of myocardial fibrils may disappear as a consequence of cellular infiltration, causing the originally heterogeneous medium of the myocardium to become more homogeneous. The change in perfusion that accompanies a rejection episode also affects transmission and reflection of ultrasound. It is known that scar or interstitial fibrosis formation follows myocyte necrosis. These changes in the myocardium may partially explain our finding. The fibrosis generates a new stiffness of elastic elements, which leads to different backscattering properties. Some researchers have conducted animal studies to investigate the use of echocardiographic tissue characterization for recognizing and differentiating between rejection and no rejection. In their published findings, all [12-141 agreed that acute rejection effects a change in the acoustic properties of myocardium that can be measured as integrated backscatter intensity or pixel intensity of echocardiograms. The results of this study suggest that the process of acute cardiac rejection is reflected in typical and reproducible changes in echocardiographic image texture. The texture becomes lighter and more heterogeneous and has greater contrast. These texture changes in the echocardiographic image can be recognized and quantified by digital image analysis. By application of Fisher’s discriminant analysis, a set of three texture parameters that describe echocardiographic image texture during cardiac rejection was selected from 117 calculated parameters. Using this set of parameters, it was possible to distinguish rejection from no rejection.
Limitations Some limitations of this study, which are inherent to both biopsy and ultrasonic evaluations, should be noted. Although endomyocardial biopsy remains the preferred method of choice for diagnosing rejection, problems with and limitations to the information gained from a myocardial specimen must be considered. A possible reason for a positive texture indicating a negative biopsy finding is the so-called sampling error, ie, an insufficient quantity of biopsy particles is taken or the tissue sample is not representative, thus causing a rejection to remain undetected. An echocardiographic evaluation window of 25 x 25 pixels corresponds approximately to a myocardial area of8 x 8mm. Aretz [15] recommended that only positive results be considered as evidence. The tissue samples should be analyzed on various sides, and the thickness of the cut should not be larger than 4 nm. He also pointed out other possible sources of false diagnostic assessments: errors caused by mixing up samples or by the formation of artifacts in the tissue sample causing a shrinkage of the myocardial area (sometimes causing it to be erroneously diagnosed as myocytolysis). Furthermore, in general, specimens should be obtained only from the upper parts of the ventricular septum (less than 5 mm underneath the
LIEBACK ET AL ECHOCARDIOGRAPHIC TISSUE CHARACTERIZATION
1167
endocardium) [16]. As a noticeable artifact of the histologic picture of the myocardial specimen, Muller and co-workers [17] described leukocytic infiltrations caused by mechanical alteration. For this reason, they recommended that the specimens be taken when the bioptome first contacts the ventricular wall. It is also known that an accumulation of inflammatory cells with focal necrosis can be the result of catecholamine effects in situations of stress or vasopressor therapy [17]. An explanation for a false-positive result using echocardiographic texture analysis is the effect of other noxae on the myocardium (eg, toxic damage or myocarditis) that change its acoustic properties. Ultrasonic tissue characterization also has pitfalls and limitations. Studies of myocardial tissue differentiation must consider cardiac dynamics (translation, rotation, and contractile motion) as well as respiratory dynamics. Because ultrasonic interaction with the heart is anisotropic, ie, varying according to the angle between ultrasonic beam and myofibrils, accurate and reproducible sound direction becomes a necessity. On the other hand, the anisotropy itself can be measured and this measurement used to present a characteristic feature of the tissue. This has not yet been tried. Further, the result of cardiac motion modifies the myocardial ”region of interest” in its depth and orientation in relation to the fixed transducer. This can lead to alterations in echo amplitudes either through changes in position of the isonified sample volume in the heart or through changes in the attenuation profile between the region of interest and the transducer. Apart from cardiac motion, the cardiac cycle itself, through its varying states of contractility, moves the acoustic indices of the myocardium [18]. For this reason, comparative examinations must be related to ECG-triggered ultrasonic images made at the same time of the cardiac cycle. Echocardiographic data collection on the heart in vivo is further limited by the relatively small acoustic window from the intercostal space. Contrary to other organs, the analyzable myocardial area is defined by relatively thin surrounding tissue, the ventricular walls. In echocardiographic tissue characterization, additional importance must be attributed to the fact that thorax and tissue situated between the transducer and the region of interest alter ultrasonic energy. When ultrasound penetrates this interposed tissue, there is a loss of energy by absorption and scattering. Apart from this, tissue has the properties of a low-pass filter, which absorbs high ultrasonic frequencies proportionally more than low ultrasonic frequencies. Thus, amplitude and character of the ultrasonic signal change when the signal passes through interposed tissue. Another major problem that influences tissue characterization is the effect of signal processing on the acoustic dates within the echocardiographic machine. To achieve reproducibility and comparability of dates between laboratories, it is necessary to develop standardized criteria for these equipment functions. Also, the range and azimuth variability of reflected ultrasound demand a central position of the region of interest in the sector.
1168
LIEBACK ET AL
Ann Thorac Surg
ECHOCARDIOGRAPHICTISSUE CHARACTERIZATION
A more fundamental problem is o u r incomplete understanding of basic mechanics in the spread of ultrasound and its interaction with tissue. Although it is possible to give a general description of the scattering processes, such a description remains oversimplified and undetailed, presenting only an approximate explanation involving one or two aspects of the processes involved. An adequate description of the reflection properties, however, should include the complex effects of scattering, diffraction, and nonlinearity. The complexity of ultrasonic processes also presents a great opportunity. Interaction of ultrasound and tissue produces an immense amount of data, thus providing information that potentially could become a major diagnostic source. Decoding and applying this information is the ultimate goal of sonographic tissue characterization.
Conclusions After taking into account the pitfalls and limitations, the following conclusions can be drawn: 1. Analysis of gray-level distribution (texture analysis) of the echocardiogram allows assessment of the structural state of the myocardium. 2 . Acute allograft rejection is reflected in a change in echocardiographic texture. 3. Using only a set of three texture parameters (inverse difference moment with the pixel distance 2 undirected, run-length nonuniformity vertical, and sector sum of the power spectrum), it is possible to distinguish rejection from no rejection. 4. Echocardiographic tissue characterization opens new noninvasive diagnostic possibilities for determining acute rejection after cardiac transplantation.
References 1. Amende I, Simon R, Seegers A, et al. Diastolic dysfunction during acute cardiac allograft rejection. Circulation 1990; Sl(Supp1 3):6&70. 2. Mimbs JW, Yuhas DE, Miller JG, Weiss AN, Sobel BE. Detection of myocardial infarction in vitro based on altered attenuation of ultrasound. Circ Res 1977;41:192-7. 3. Sagar KB, Pelc R, Rhyne SD, Howard J, Warltier DC. Estimation of myocardial infarct size with ultrasonic tissue characterization. Circulation 1991;83:1419-28. 4. Lieback E, Meyer R, Romaniuk P, et al. Ultrasonic diagnosis of myocarditis using texture analysis of two dimensional echocardiograms. Z Gesamte Inn Med 1989;44:484-7. 5. Fraker TD, Bingle JF, Wilkerson RD, Klingler JW, Weaver MT. Acute myocardial ischemia detected in dogs by temporal variation in two-dimensional ultrasound gray level. Am Heart J 1988;116:249-53. 6. Haendchen RV, Ong K, Fishbein MC. Early differentiation of infarcted and noninfarcted reperfused myocardium in dogs by quantitative analysis of regional myocardial echo amplitudes. Circ Res 1985;57:71&28. 7. Sakabe Y, Hishida H, Kawamura K, et al. Ultrasonic tissue characterization in diagnosing myocardial infarction. J Cardiogr 1987;17:49-58. 8. Ferdighini EM, Pinamonti 8, Picano E, et al. Quantitative texture analysis in echocardiography: application to the diagnosis of myocarditis. J Clin Ultrasound 1991;19:263-70. 9. Billingham ME. Diagnosis of cardiac rejection by endomyocardial biopsy. Heart Transplant 1981;1:2530.
1994;57116&70
10. Dawkins KD, Oldershaw PJ, Billingham ME, et al. Changes in diastolic function as a noninvasive marker of cardiac allograft rejection. Heart Transplant 1984;3:286-92. 11. Valantine HA, Fower MB, Hunt SA. Changes in Doppler echocardiographic indices of left ventricular function a s potential markers of acute rejection. Circulation 1986;74:159-75. 2. Chandrasekaran K, Bansal RC, Greenleaf JF, et al. Early recognition of heart transplant rejection by backscatter analysis from serial 2D echos in a heterotopic transplant mo'del. J Heart Transplant 1987;6:1-7. 3. Dawkins KD, Haverich A, Aziz S, Billingham ME, Jamieson SW, Gibson DG. Detection of acute cardiac rejection using color echocardiography. Circulation 1985;72(Suppl3):207. 4. Masuyama T, Valantine HA, Gibbons R, Schnittger I, Popp RL. Serial measurement of integrated ultrasonic backscatter in human cardiac allografts for the recognition of acute rejection. Circulation 1990;81:829-39. 15. Aretz HT. Myocarditis: the Dallas criteria. Hum Pathol 1987; 18:619-23. 16. Herskowitz A, Soule LM, Mellits ED. Histologic predictors of acute cardiac rejection in human endomyocardial biopsies: a multivariate analysis. J Am Coll Cardiol 1987;9:802-10. 17. Muller S, Muller P, Meyer R, et al. Modern concepts in the diagnosis of myocarditis. Eur Heart J 1987;8:35-8. 18. Olshansky B, Collins SM, Skorton DJ, Prasad NV. Alteration of left ventricular wall gray level in two-dimensional echocardiograms due to cardiac contraction. Circulation 1984;70: 972-7. 19. Haralick RM. Statical and structural approaches to texture. Proc IEEE 1979;67:785-804. 20. Galloway MM. Texture analysis using gray level run lengths. Comp Graph Imag Process 1975;4:172-5. 21. Weszka JS, Dyer CR, Rosenfeld A. A comparative study of texture measures for terrain classification. IEEE Trans Sys Man Cybern 1976;6:269-85.
Appendix 1. I.
Texture Analysis
A. First-order histogram First-order texture is a type of spatial gray-level distribution identifiable by its gray-level histogram that indicates the frequencies of gray levels present in the image. Parameters derived from the first-order histogram describe the shape of the first-order gray-level distribution without considering spatial interdependencies.
1. Mean gray level:
cc 1
Gave = (UN)
1
g(i, j)
1=1j=1
where g (i, j) = gray level of pixel (i, j); g (i, j) = O..G; i = 1..I; j = 1..J; G = maximum gray level of the image; 0 = minimum gray level of the image; I = number of pixels in an image line; J = number of pixels in an image column; and N = number of pixels in the image. 2. Variance:
3. Skewness:
LIEBACK ET AL ECHOCARDIOGRAPHICTISSUE CHARACTERIZATION
Ann Thorac Surg 1994;571164-70
1169
1. Short-run emphasis (SRE):
4. Access:
p (i, j) = (i, j)th entry in the given run-length matrix; N, = number of gray levels in the image; N, = number of different run lengths; and P = number of pixels in the image.
5. Percentile:
‘5‘H(g) < p/100N
=
<
g=o
gP
C k(g)
g=o
Because of the lack of spatial information, these parameters were not able to precisely quantify the number, size, and orientation of localized texture structures. Therefore, when texture features like contrast, heterogeneity, and local inhomogeneities are of interest, second-order statistics should be used for analysis.
B. Co-occurrence matrix The co-occurrence matrix is a two-dimensional histogram that characterizes the occurrence of gray-level combinations in pairs of spatially related pixels. It can be calculated along different axes (west-east, north-south, northwest-southeast, northeast-southwest, and undirected) and between image points separated by a variable number of pixels. For this study, the co-occurrence matrix was calculated at distances of 1, 2, 4, and 6 pixels along each axis. Haralick [19] suggested a group of texture parameters that can be derived from a co-occurrence matrix: 1. Second difference moment (SDM):
SDM = c c ( k - I)’ k
X
c(k, 1)
l
where c (k, 1) = (k, 1)th entry in the normalized co-occurrence matrix; k = O..G; I = O..G; G = maximum gray level of the image; and 0 = minimum gray level of the image.
2. Entropy (ENT): ENT =
-
2. Long-run emphasis (LRE):
3. Gray-level nonuniformity (GLN):
i = l j=l
4. Run-length nonuniformity (RLN):
5. Run percentage (RP):
i=l j=1
The run-length statistic primarily describes the heterogeneity of gray values and the relative size of specific echocardiographic reflections in the region of interest [20].
D. Power spectrum The two-dimensional power spectrum is calculated from the cosinus transformation /F/ of the digital image. The power spectrum is then /FP. The cosinus transformation of the image g (i, j) is defined by:
C C c ( k , 1) X log c(k, 1) k l
3. Angular second moment (ASM): ASM = Z C c ( k , 1)’ k l
4. Inverse difference moment (IDM):
IDM = CC k I
c(k, 1) 1 + (k - 1)’
wherei = 0, I - 1; j = 0, J - 1;u = 0, I - 1;v = 0, J - I;I = number of pixels in a line of a digitized image; and J = number of pixels in a column of a digitized image. It is known that the ring sum of the values of /F/’ is a parameter for coarseness of texture. The ring sums were calculated by:
C. Run-Length Analysis Gray-level run-length data count the number of gray-level runs by their length and gray-level range. A gray-level run is a set of neighboring pixels having the same gray level. The length of the run is the number of image points it contains. A gray-level run-length matrix, in which each matrix element p(i, j) specified the number of runs of gray level i and length j, was calculated for each image. It was computed along both horizontal and vertical axes. To obtain numerical measurements from the matrices, one can compute functions analogous to those used by Haralick [19] for co-occurrence matrix. In the present study, five run-length features were calculated from the matrix:
C
QrII2=
2
IF(u, 41’
r12 5 “2 t <1 2 osu,v5n-1
Using these parameters, homogeneity, heterogeneity, and contrast of a texture can be described.
The sector sum of values of the power spectrum describes the directional pattern of texture [21]. The sector sums were calculated by: Qs,o,
=
C
IF(u, v)1’
8 , s tan-’(v/u) < 02
O5u.vrn-1
11. Data Analysis and Classification
Parameters can be selected by downward processing, upward processing, and a combination of both processes. Using the upward process and adding a further criterion tested whether discrimination can be improved. The selection of an “optimal” combination of criteria from the total amount of all
1170
LIEBACK ET AL ECHOCARDIOGRAPHIC TISSUE CHARACTERIZATION
selected classification criteria was, in general, guided by the consideration of estimated risk factors (estimation of erroneous classification). For selecting criteria of texture parameters, the assessment of Bayes’s risk and a measurement of distance was used. Bayes’s risk estimates to what extent a classification principle is correct. In case of equal loss in each of the classification groups, it is identical with the estimation of erroneous classification. In the downward process, the F test was applied to all texture parameters to check whether the distance between the distribution of specific classification groups had become less (indispensability) after elimination of a specific classification criterion (texture parameter). “Separa-
Ann Thorac Surg 1994;57116470
bility” loss in the absence of a criterion was greater the smaller the level of significance. The combination of upward processing (increase in amount of parameters) and downward processing (reduction in amount of parameters) was necessary because the time required to calculate data was so great, an evaluation of all parameter combinations was not possible. Therefore it was not certain whether after discontinuation of the discriminant selection process, a still smaller parameter combination with equally good separability properties might exist. A combination of upward and downward processing was used for evaluation in this study for this reason.
Important Notice From the Southern Thoracic Surgical Association The March 1994 issue of The Annals carried an incorrect announcement about the meeting dates for the Southern Thoracic Surgical Association. The correct dates for the Forty-first Annual Meeting of the Southern Thoracic Surgical Association to be held at Marriott’s Marco Island Resort and Golf Club, Marco Island, Florida, are NOVEMBER 1612, 1994. The Postgraduate Course will be held the morning of Thursday, NOVEMBER 10, 1994, and will provide in-depth coverage of thoracic surgical topics selected primarily as a means to enhance and broaden the knowledge of practicing thoracic and cardiac surgeons. Members wishing to participate in the Scientific Program should submit an original abstract and one copy by May 15, 1994, to Kamal A. Mansour, MD, Program Chairman, Southern Thoracic Surgical Association, 401 North Michigan Avenue, Chicago, IL 606114267. Abstracts must be submitted on the Southern Thoracic Sur-
gical Association abstract form. These forms may be obtained from the Association’s office or in the April issue of The Annals of Thoracic Surgery. Manuscripts of accepted papers must be submitted to The Annals of Thoracic Surgery prior to the 1994 meeting or to the Secretary-Treasurer at the opening of the Scientific Session. Applications for membership should be completed by August 1, 1994, and forwarded to John M. Kratz, MD, Membership Committee Chairman, Southern Thoracic Surgical Association, 401 North Michigan Avenue, Chicago, IL 60611-4267.
Hendrick B. Barner, M D Secretary-Treasurer Southern Thoracic Surgical Association 401 North Michigan Avenue Chicago, IL 60611-4267
Ultrasonic tissue characterization is based on the assumption that microscopic tissue structures are identifiable by their acoustic properties. Our study group consisted of 23 cardiac recipients. Two-dimensional images were obtained within 2 hours of endomyocardial biopsy. The end-diastolic echo frames were digitized into the matrix of an image-processing system. A region of interest was placed into the anteroseptal segment of the left ventricle. The texture within the region of interest was analyzed using four major groups of texture analysis (first-qrder histogram, co-occurrence matrix, run-length statistic, and power spectrum). A total of 408 echocardiographic examinations were compared with histologic
findings. The 117 initially calculated texture parameters were reduced incrementally using a series of discriminant analyses. A set of three texture parameters (inverse difference moment undirected, run-length nonuniformity vertical, and sector sum) was able to describe changed echocardiographic texture when rejection occurred. Using these three parameters, echocardiographic sensitivity was 89.0% and specificity was 83.6% for moderate rejection. We conclude that cardiac rejection is associated with echocardiographic texture alterations and that serial echocardiographic texture analysis can reliably identify rejection. (Ann Thorac Surg 1994;57:1168&70)
T
Postoperative Management
o date, echocardiographic diagnosis of cardiac rejection has focused mainly on the analysis of functional, particularly diastolic, parameters of the heart [l]. The purpose of this study was to evaluate the ability of texture analysis of echocardiograms to detect morphologic changes in the myocardium that were caused by rejection after cardiac transplantation. This research is based on the assumption that abnormalities in microscopic tissue structure are reflected in its acoustic properties. Current approaches to ultrasonic tissue characterization include measurement of acoustic parameters such as attenuation coefficients [2] and integrated backscatter [3] and digital image analysis of two-dimensional echocardiograms [4]. Ultrasonic tissue characterization research has shown that these techniques can identify acute ischemic [5], reperfused [6], infarcted [7], and inflamed [8] myocardium. This study examined whether rejection is associated with an abnormal pattern of echocardiographic texture and whether the evolution of the disease can be recognized by means of quantitative texture analysis.
Material and Methods The study group consisted of 23 patients (3 women and 20 men) who had had orthotopic cardiac transplantation. Mean age was 37.2 & 9.0 years (range, 19 to 48 years). Follow-up ranged from 3 to 16 months. Accepted for publication Aug 12, 1993. Address reprint requests to Dr Lieback, German Heart Institute, Augustenburger Platz 1, 13353 Berlin, Germany.
0
1994 by The Society of Thoracic Surgeons
All patients received standard triple-drug immunosuppression therapy with cyclosporine, azathioprine, and prednisone. Cytolytic therapy with antithymocyte globulin was given on postoperative days 1 to 3. Cardiac rejection was monitored using endomyocardial biopsy performed weekly during the first postoperative month and biweekly during the second and third months. Biopsy intervals became longer later in the postoperative course. Biopsy specimens were graded according to the classification of Billingham [9]. Moderate and severe episodes of rejection were treated with pulsed methylprednisolone (0.5 g/day for 3 days). In the case of persistent rejection, rabbit antithymocyte globulin, or OKT3 if necessary, was administered over another 3 days.
Standardized Echocardiographic Examination and linage Digitization Echocardiographic examinations were performed within 2 hours of endomyocardial biopsy without knowledge of the histologic result. The examinations were carefully standardized. Two-dimensional echocardiograms were produced using an electronic 90-degree sector-scanning system operating at 3.75 MHz. Short-axis images were obtained at the midventricular papillary muscle Level. Special care was taken that the angle of the interrogating ultrasound wave in the anteroseptal segment was approximately 90 degrees. Amplification adjustment was standardized for each patient, ie, the first amplification value used on each patient was maintained throughout the study. In general, amplifications between 72 and 76 dB were used. The other variables, such as dynamic range, enhancement, reject, and gamma correction, were kept 0003-4975/94/$7.00
LIEBACK ET AL ECHOCARDIOGRAPHICTISSUE CHARACTERIZATION
Ann Thorac Surg 1994;571164-70
ZEE
I
1165
occurrence matrix, run-length analysis, and power spectrum (Appendix 1).
Intensity
Data Analysis and Classification
Pixel
A
25
Intensity
25
Analysis of ultrasonic texture data is performed to determine which combination of texture variables is necessary and adequate for distinguishing between various histologic conditions of the myocardium (rejection versus no rejection) after cardiac transplantation. The 117 initially calculated texture parameters were reduced to an adequate number using a series of discriminant analyses. This parameter selection process was done to obtain a number of parameters that was relatively low and in which erroneous classification was relatively rare (see Appendix 1). After calculating the classification principle ("optimal" texture parameter set), the computer classified the data from the texture analysis of the 408 echocardiograms obtained during the study. Sensitivity, specificity, and accuracy of texture analysis in comparison with the endomyocardial biopsy results were calculated according to the following formulas: Sensitivity = TP/(TP + FN) Specificity = TN/(FP + TN)
Pixel
25
B Fig 1. Example of three-dimensional pixel maps of gray-level distribution in the region of interest of a representative patient in the phase without rejection and at the time of rejection. Note (A) the relatively homogeneous gray-level values in the phase without rejection and ( B ) the high differences in gray-level values at the time of rejection.
constant. Depth control was standardized at 15 cm. During the study, a total of 408 echocardiographic examinations were conducted. The electrocardiogram-triggered end-diastolic images were digitized into the imageprocessing system.
Endomyocardial Biopsy A median of five tissue samples were taken by endomyocardial biopsy. Biopsy specimens were considered positive when two independent examiners arrived at the same results indicating either moderate or severe rejection.
Texture Analysis The pattern of gray-level distribution in an image is called texture. Echocardiographic texture is defined as regional distribution of echo amplitudes, which are projected on the screen as gray levels. Regions of interest were chosen by using an interactive computer program. These areas, having a standard size of 25 x 25 pixels, were placed into the anteroseptal segment of the left ventricle. The texture within the region of interest was analyzed using measurements related to intensity and distribution of echo amplitudes (gray levels). In the present study, four major types of texture analysis were used: first-order histogram, co-
Accuracy = (TP + TN)/(TP + FN + FP + TN), where TP = true-positive, TN = true-negative, FP false-positive, and FN = false-negative.
=
Results Echocardiographic texture of the myocardium changed significantly when rejection began to develop. Figure 1 shows the gray-level distribution in the region of interest of a representative patient before rejection and during rejection. For quantification of gray-level distribution, 117 texture parameters were calculated. The first 131 echocardiographicexaminationsand the corresponding endomyocardial biopsy specimens were used as the learning group. Analysis of the myocardial biopsy samples produced the following results: no rejection, 54; rejection, 39; and fibrosis, 38. The 117 texture parameters were incrementally reduced by means of upward and downward processing of discriminant analysis. The following parameters were found sufficient to effectively distinguish between rejection, fibrosis, and no rejection: sector sum 0 degrees to 30 degrees; run-length nonuniformity vertical; inverse differ-
+ ++ + ,
4.38
L++++%+ M 1
B I
7.77
OD
M r
3
0:
* 11.2
* **
o&X* ,
M
*$*
8
2 14.6
Fig 2. Distribution of the first nonelementary discriminant function of the learning class showing the ability to distinguish between "rejec= l mean value; + = rejection; * = no tion'' and "no rejection." & rejection; 0 = fibrosis.)
1166
LIEBACK ET AL ECHOCARDIOGRAPHIC TISSUE CHARACTERIZATION
ence moment (dx = 2 undirected). The distribution of the observations concerning the first nonelementary discriminant functions and their classification in no rejection, rejection, and fibrosis is shown in Figure 2. The risk of classification according to Bayes amounted to 0.108425. After the classification principle (texture parameter set) had been found by discriminant analysis, the parameter set was used to classify all 408 data sets of texture analysis. These data sets were compared with their corresponding histologic examination data sets for variances in sensitivity, specificity, and accuracy. Results of no rejection and fibrosis were added to indicate "biopsy negative." The results are presented in Table 1. The sensitivity of the computer classification was 89% and the specificity was 83.6%. Single-case analyses demonstrated that texture parameters can indicate rejection early. Therefore, it is very important to note the changes in echocardiographic texture after cardiac transplantation. The postoperative course of a selected patient is presented in Figure 3. In this patient, the first three biopsy controls were negative, and echocardiographic texture at this point was normal. On the 31st postoperative day, echocardiographicimage texture began changing, and among the changes was a decrease in run-length nonuniformity. The biopsy specimen did not yet indicate rejection. However, a subsequent biopsy specimen 7 days later showed signs of moderate rejection. The run-length uniformity continued to be low. Immunosuppression therapy was applied. After 2 weeks, further examination revealed normal histologic conditions. Echocardiographic texture also returned to normal. During the later postoperative course, texture parameter changes coincided with biopsy findings of rejection.
Comment In previous attempts to diagnose cardiac rejection echocardiographically, quantitative analysis of the effects of rejection-induced changes in organic function was aimed at factors such as increased ventricular wall thickening and left ventricular muscle volume. These changes indicated early manifestation of myocardial edema in connection with rejection before the introduction of cyclosporine. When therapy was combined with cyclosporine, myocardial edema became less pronounced. Because patients receiving cyclosporine often had development of
Table 1. Computer Classification of "Rejection" and "No Reiection'" Computer Diagnosis Biopsy Diagnosis Rejection No rejection
Rejection
No Rejection
75 32
9 272
Using a set of three texture parameters (inverse difference moment, run-length nonuniformity, and sector sum), the echocardiographic sensitivity for rejection was 89.0%, the specificity was 83.6%, and the accuracy was 85.0%.
a
Ann Thorac Surg
1994;57116470
Run-length nonunilorrnlly
80 75 70 65 60 55
50 45 40 35 30 25 20 15 10
5
n 7
BIOPSY +o
TREATMENT
31
38
52
3 tk t 0
no
14
21
moderate
***
A
3 x 0.5 aMP
08
A
no
88
108 115
A
0
no A moderatgo
*.*
3 x 0.5 aMP
135
A
no
102 170 days ~ o . 1 HTx
0' **.
3 I 0.5 a M P
Fig 3. Postoperative course of a representative patient. The first three biopsy controls were negative, and echocardiographic texture was; normal. On the 31st day, the run-length nonuniformity decreased, but endomyocardial biopsy sample did not indicate rejection. The biopsy control 7 days later show signs of moderate rejection. Run-length nonuniformity continued to be low. After immunosuppression therapy, an examination revealed normal histologic conditions. Echocardiogrcrphic texture also returned to normal. (HTx = cardiac transplantation; MP = methylprednisolone.)
arterial hypertension with early left ventricular hypertrophy, wall thickening or muscle volume, which is deduced from it, began to lose diagnostic value. Other researchers have examined to what extent acute rejection is able to cause changes in systolic and diastolic cardiac function. During the rejection process, few or no changes in systolic function were found. However, left ventricular diastolic dynamics are affected by alterations in isovolumetric relaxation, early diastolic filling, anld left ventricular distensibility [lo, 111. Impaired diastolic cardiac function can be explained as a result of increased stiffness through interstitial cellular infiltration and myocardial edema, thus leading to a reduction in compliance of the left ventricle. Other researchers occasionally encountered false-negative results, even in cases of severe rejection, thus making assessment of diastolic ventricular function alone inadequate for reliable diagnosis. In this study, acoustic properties of cardiac allografts were assessed by measuring the texture parameters of echocardiograms. Although the relationship between echocardiographic image texture and anatomic composition of tissue is complex and not yet completely uinderstood, one can assume that the processes connected with rejection, such as cellular infiltration, myocytolysis, linterstitial edema, and changed perfusion, influence the acoustic properties of the myocardium and, with this, the echocardiographic image texture. The distance between the scatters is enlarged by interstitial edema, ie. the number of scatters per volume unit diminishes. The increased liquid content of the tissue, on the other hand, influences attenuation qualities, ie, increased water content decreases attenuation. The density and elasticity of the myocardium are changed by cellular infiltration, a
Ann Thorac Surg 1994;571164-70
condition usually causing increased reflection. At the same time, the size and shape of scatters change. Changes in fiber orientation and rearrangement of dominant myofibrils occur. The edges of myocardial fibrils may disappear as a consequence of cellular infiltration, causing the originally heterogeneous medium of the myocardium to become more homogeneous. The change in perfusion that accompanies a rejection episode also affects transmission and reflection of ultrasound. It is known that scar or interstitial fibrosis formation follows myocyte necrosis. These changes in the myocardium may partially explain our finding. The fibrosis generates a new stiffness of elastic elements, which leads to different backscattering properties. Some researchers have conducted animal studies to investigate the use of echocardiographic tissue characterization for recognizing and differentiating between rejection and no rejection. In their published findings, all [12-141 agreed that acute rejection effects a change in the acoustic properties of myocardium that can be measured as integrated backscatter intensity or pixel intensity of echocardiograms. The results of this study suggest that the process of acute cardiac rejection is reflected in typical and reproducible changes in echocardiographic image texture. The texture becomes lighter and more heterogeneous and has greater contrast. These texture changes in the echocardiographic image can be recognized and quantified by digital image analysis. By application of Fisher’s discriminant analysis, a set of three texture parameters that describe echocardiographic image texture during cardiac rejection was selected from 117 calculated parameters. Using this set of parameters, it was possible to distinguish rejection from no rejection.
Limitations Some limitations of this study, which are inherent to both biopsy and ultrasonic evaluations, should be noted. Although endomyocardial biopsy remains the preferred method of choice for diagnosing rejection, problems with and limitations to the information gained from a myocardial specimen must be considered. A possible reason for a positive texture indicating a negative biopsy finding is the so-called sampling error, ie, an insufficient quantity of biopsy particles is taken or the tissue sample is not representative, thus causing a rejection to remain undetected. An echocardiographic evaluation window of 25 x 25 pixels corresponds approximately to a myocardial area of8 x 8mm. Aretz [15] recommended that only positive results be considered as evidence. The tissue samples should be analyzed on various sides, and the thickness of the cut should not be larger than 4 nm. He also pointed out other possible sources of false diagnostic assessments: errors caused by mixing up samples or by the formation of artifacts in the tissue sample causing a shrinkage of the myocardial area (sometimes causing it to be erroneously diagnosed as myocytolysis). Furthermore, in general, specimens should be obtained only from the upper parts of the ventricular septum (less than 5 mm underneath the
LIEBACK ET AL ECHOCARDIOGRAPHIC TISSUE CHARACTERIZATION
1167
endocardium) [16]. As a noticeable artifact of the histologic picture of the myocardial specimen, Muller and co-workers [17] described leukocytic infiltrations caused by mechanical alteration. For this reason, they recommended that the specimens be taken when the bioptome first contacts the ventricular wall. It is also known that an accumulation of inflammatory cells with focal necrosis can be the result of catecholamine effects in situations of stress or vasopressor therapy [17]. An explanation for a false-positive result using echocardiographic texture analysis is the effect of other noxae on the myocardium (eg, toxic damage or myocarditis) that change its acoustic properties. Ultrasonic tissue characterization also has pitfalls and limitations. Studies of myocardial tissue differentiation must consider cardiac dynamics (translation, rotation, and contractile motion) as well as respiratory dynamics. Because ultrasonic interaction with the heart is anisotropic, ie, varying according to the angle between ultrasonic beam and myofibrils, accurate and reproducible sound direction becomes a necessity. On the other hand, the anisotropy itself can be measured and this measurement used to present a characteristic feature of the tissue. This has not yet been tried. Further, the result of cardiac motion modifies the myocardial ”region of interest” in its depth and orientation in relation to the fixed transducer. This can lead to alterations in echo amplitudes either through changes in position of the isonified sample volume in the heart or through changes in the attenuation profile between the region of interest and the transducer. Apart from cardiac motion, the cardiac cycle itself, through its varying states of contractility, moves the acoustic indices of the myocardium [18]. For this reason, comparative examinations must be related to ECG-triggered ultrasonic images made at the same time of the cardiac cycle. Echocardiographic data collection on the heart in vivo is further limited by the relatively small acoustic window from the intercostal space. Contrary to other organs, the analyzable myocardial area is defined by relatively thin surrounding tissue, the ventricular walls. In echocardiographic tissue characterization, additional importance must be attributed to the fact that thorax and tissue situated between the transducer and the region of interest alter ultrasonic energy. When ultrasound penetrates this interposed tissue, there is a loss of energy by absorption and scattering. Apart from this, tissue has the properties of a low-pass filter, which absorbs high ultrasonic frequencies proportionally more than low ultrasonic frequencies. Thus, amplitude and character of the ultrasonic signal change when the signal passes through interposed tissue. Another major problem that influences tissue characterization is the effect of signal processing on the acoustic dates within the echocardiographic machine. To achieve reproducibility and comparability of dates between laboratories, it is necessary to develop standardized criteria for these equipment functions. Also, the range and azimuth variability of reflected ultrasound demand a central position of the region of interest in the sector.
1168
LIEBACK ET AL
Ann Thorac Surg
ECHOCARDIOGRAPHICTISSUE CHARACTERIZATION
A more fundamental problem is o u r incomplete understanding of basic mechanics in the spread of ultrasound and its interaction with tissue. Although it is possible to give a general description of the scattering processes, such a description remains oversimplified and undetailed, presenting only an approximate explanation involving one or two aspects of the processes involved. An adequate description of the reflection properties, however, should include the complex effects of scattering, diffraction, and nonlinearity. The complexity of ultrasonic processes also presents a great opportunity. Interaction of ultrasound and tissue produces an immense amount of data, thus providing information that potentially could become a major diagnostic source. Decoding and applying this information is the ultimate goal of sonographic tissue characterization.
Conclusions After taking into account the pitfalls and limitations, the following conclusions can be drawn: 1. Analysis of gray-level distribution (texture analysis) of the echocardiogram allows assessment of the structural state of the myocardium. 2 . Acute allograft rejection is reflected in a change in echocardiographic texture. 3. Using only a set of three texture parameters (inverse difference moment with the pixel distance 2 undirected, run-length nonuniformity vertical, and sector sum of the power spectrum), it is possible to distinguish rejection from no rejection. 4. Echocardiographic tissue characterization opens new noninvasive diagnostic possibilities for determining acute rejection after cardiac transplantation.
References 1. Amende I, Simon R, Seegers A, et al. Diastolic dysfunction during acute cardiac allograft rejection. Circulation 1990; Sl(Supp1 3):6&70. 2. Mimbs JW, Yuhas DE, Miller JG, Weiss AN, Sobel BE. Detection of myocardial infarction in vitro based on altered attenuation of ultrasound. Circ Res 1977;41:192-7. 3. Sagar KB, Pelc R, Rhyne SD, Howard J, Warltier DC. Estimation of myocardial infarct size with ultrasonic tissue characterization. Circulation 1991;83:1419-28. 4. Lieback E, Meyer R, Romaniuk P, et al. Ultrasonic diagnosis of myocarditis using texture analysis of two dimensional echocardiograms. Z Gesamte Inn Med 1989;44:484-7. 5. Fraker TD, Bingle JF, Wilkerson RD, Klingler JW, Weaver MT. Acute myocardial ischemia detected in dogs by temporal variation in two-dimensional ultrasound gray level. Am Heart J 1988;116:249-53. 6. Haendchen RV, Ong K, Fishbein MC. Early differentiation of infarcted and noninfarcted reperfused myocardium in dogs by quantitative analysis of regional myocardial echo amplitudes. Circ Res 1985;57:71&28. 7. Sakabe Y, Hishida H, Kawamura K, et al. Ultrasonic tissue characterization in diagnosing myocardial infarction. J Cardiogr 1987;17:49-58. 8. Ferdighini EM, Pinamonti 8, Picano E, et al. Quantitative texture analysis in echocardiography: application to the diagnosis of myocarditis. J Clin Ultrasound 1991;19:263-70. 9. Billingham ME. Diagnosis of cardiac rejection by endomyocardial biopsy. Heart Transplant 1981;1:2530.
1994;57116&70
10. Dawkins KD, Oldershaw PJ, Billingham ME, et al. Changes in diastolic function as a noninvasive marker of cardiac allograft rejection. Heart Transplant 1984;3:286-92. 11. Valantine HA, Fower MB, Hunt SA. Changes in Doppler echocardiographic indices of left ventricular function a s potential markers of acute rejection. Circulation 1986;74:159-75. 2. Chandrasekaran K, Bansal RC, Greenleaf JF, et al. Early recognition of heart transplant rejection by backscatter analysis from serial 2D echos in a heterotopic transplant mo'del. J Heart Transplant 1987;6:1-7. 3. Dawkins KD, Haverich A, Aziz S, Billingham ME, Jamieson SW, Gibson DG. Detection of acute cardiac rejection using color echocardiography. Circulation 1985;72(Suppl3):207. 4. Masuyama T, Valantine HA, Gibbons R, Schnittger I, Popp RL. Serial measurement of integrated ultrasonic backscatter in human cardiac allografts for the recognition of acute rejection. Circulation 1990;81:829-39. 15. Aretz HT. Myocarditis: the Dallas criteria. Hum Pathol 1987; 18:619-23. 16. Herskowitz A, Soule LM, Mellits ED. Histologic predictors of acute cardiac rejection in human endomyocardial biopsies: a multivariate analysis. J Am Coll Cardiol 1987;9:802-10. 17. Muller S, Muller P, Meyer R, et al. Modern concepts in the diagnosis of myocarditis. Eur Heart J 1987;8:35-8. 18. Olshansky B, Collins SM, Skorton DJ, Prasad NV. Alteration of left ventricular wall gray level in two-dimensional echocardiograms due to cardiac contraction. Circulation 1984;70: 972-7. 19. Haralick RM. Statical and structural approaches to texture. Proc IEEE 1979;67:785-804. 20. Galloway MM. Texture analysis using gray level run lengths. Comp Graph Imag Process 1975;4:172-5. 21. Weszka JS, Dyer CR, Rosenfeld A. A comparative study of texture measures for terrain classification. IEEE Trans Sys Man Cybern 1976;6:269-85.
Appendix 1. I.
Texture Analysis
A. First-order histogram First-order texture is a type of spatial gray-level distribution identifiable by its gray-level histogram that indicates the frequencies of gray levels present in the image. Parameters derived from the first-order histogram describe the shape of the first-order gray-level distribution without considering spatial interdependencies.
1. Mean gray level:
cc 1
Gave = (UN)
1
g(i, j)
1=1j=1
where g (i, j) = gray level of pixel (i, j); g (i, j) = O..G; i = 1..I; j = 1..J; G = maximum gray level of the image; 0 = minimum gray level of the image; I = number of pixels in an image line; J = number of pixels in an image column; and N = number of pixels in the image. 2. Variance:
3. Skewness:
LIEBACK ET AL ECHOCARDIOGRAPHICTISSUE CHARACTERIZATION
Ann Thorac Surg 1994;571164-70
1169
1. Short-run emphasis (SRE):
4. Access:
p (i, j) = (i, j)th entry in the given run-length matrix; N, = number of gray levels in the image; N, = number of different run lengths; and P = number of pixels in the image.
5. Percentile:
‘5‘H(g) < p/100N
=
<
g=o
gP
C k(g)
g=o
Because of the lack of spatial information, these parameters were not able to precisely quantify the number, size, and orientation of localized texture structures. Therefore, when texture features like contrast, heterogeneity, and local inhomogeneities are of interest, second-order statistics should be used for analysis.
B. Co-occurrence matrix The co-occurrence matrix is a two-dimensional histogram that characterizes the occurrence of gray-level combinations in pairs of spatially related pixels. It can be calculated along different axes (west-east, north-south, northwest-southeast, northeast-southwest, and undirected) and between image points separated by a variable number of pixels. For this study, the co-occurrence matrix was calculated at distances of 1, 2, 4, and 6 pixels along each axis. Haralick [19] suggested a group of texture parameters that can be derived from a co-occurrence matrix: 1. Second difference moment (SDM):
SDM = c c ( k - I)’ k
X
c(k, 1)
l
where c (k, 1) = (k, 1)th entry in the normalized co-occurrence matrix; k = O..G; I = O..G; G = maximum gray level of the image; and 0 = minimum gray level of the image.
2. Entropy (ENT): ENT =
-
2. Long-run emphasis (LRE):
3. Gray-level nonuniformity (GLN):
i = l j=l
4. Run-length nonuniformity (RLN):
5. Run percentage (RP):
i=l j=1
The run-length statistic primarily describes the heterogeneity of gray values and the relative size of specific echocardiographic reflections in the region of interest [20].
D. Power spectrum The two-dimensional power spectrum is calculated from the cosinus transformation /F/ of the digital image. The power spectrum is then /FP. The cosinus transformation of the image g (i, j) is defined by:
C C c ( k , 1) X log c(k, 1) k l
3. Angular second moment (ASM): ASM = Z C c ( k , 1)’ k l
4. Inverse difference moment (IDM):
IDM = CC k I
c(k, 1) 1 + (k - 1)’
wherei = 0, I - 1; j = 0, J - 1;u = 0, I - 1;v = 0, J - I;I = number of pixels in a line of a digitized image; and J = number of pixels in a column of a digitized image. It is known that the ring sum of the values of /F/’ is a parameter for coarseness of texture. The ring sums were calculated by:
C. Run-Length Analysis Gray-level run-length data count the number of gray-level runs by their length and gray-level range. A gray-level run is a set of neighboring pixels having the same gray level. The length of the run is the number of image points it contains. A gray-level run-length matrix, in which each matrix element p(i, j) specified the number of runs of gray level i and length j, was calculated for each image. It was computed along both horizontal and vertical axes. To obtain numerical measurements from the matrices, one can compute functions analogous to those used by Haralick [19] for co-occurrence matrix. In the present study, five run-length features were calculated from the matrix:
C
QrII2=
2
IF(u, 41’
r12 5 “2 t <1 2 osu,v5n-1
Using these parameters, homogeneity, heterogeneity, and contrast of a texture can be described.
The sector sum of values of the power spectrum describes the directional pattern of texture [21]. The sector sums were calculated by: Qs,o,
=
C
IF(u, v)1’
8 , s tan-’(v/u) < 02
O5u.vrn-1
11. Data Analysis and Classification
Parameters can be selected by downward processing, upward processing, and a combination of both processes. Using the upward process and adding a further criterion tested whether discrimination can be improved. The selection of an “optimal” combination of criteria from the total amount of all
1170
LIEBACK ET AL ECHOCARDIOGRAPHIC TISSUE CHARACTERIZATION
selected classification criteria was, in general, guided by the consideration of estimated risk factors (estimation of erroneous classification). For selecting criteria of texture parameters, the assessment of Bayes’s risk and a measurement of distance was used. Bayes’s risk estimates to what extent a classification principle is correct. In case of equal loss in each of the classification groups, it is identical with the estimation of erroneous classification. In the downward process, the F test was applied to all texture parameters to check whether the distance between the distribution of specific classification groups had become less (indispensability) after elimination of a specific classification criterion (texture parameter). “Separa-
Ann Thorac Surg 1994;57116470
bility” loss in the absence of a criterion was greater the smaller the level of significance. The combination of upward processing (increase in amount of parameters) and downward processing (reduction in amount of parameters) was necessary because the time required to calculate data was so great, an evaluation of all parameter combinations was not possible. Therefore it was not certain whether after discontinuation of the discriminant selection process, a still smaller parameter combination with equally good separability properties might exist. A combination of upward and downward processing was used for evaluation in this study for this reason.
Important Notice From the Southern Thoracic Surgical Association The March 1994 issue of The Annals carried an incorrect announcement about the meeting dates for the Southern Thoracic Surgical Association. The correct dates for the Forty-first Annual Meeting of the Southern Thoracic Surgical Association to be held at Marriott’s Marco Island Resort and Golf Club, Marco Island, Florida, are NOVEMBER 1612, 1994. The Postgraduate Course will be held the morning of Thursday, NOVEMBER 10, 1994, and will provide in-depth coverage of thoracic surgical topics selected primarily as a means to enhance and broaden the knowledge of practicing thoracic and cardiac surgeons. Members wishing to participate in the Scientific Program should submit an original abstract and one copy by May 15, 1994, to Kamal A. Mansour, MD, Program Chairman, Southern Thoracic Surgical Association, 401 North Michigan Avenue, Chicago, IL 606114267. Abstracts must be submitted on the Southern Thoracic Sur-
gical Association abstract form. These forms may be obtained from the Association’s office or in the April issue of The Annals of Thoracic Surgery. Manuscripts of accepted papers must be submitted to The Annals of Thoracic Surgery prior to the 1994 meeting or to the Secretary-Treasurer at the opening of the Scientific Session. Applications for membership should be completed by August 1, 1994, and forwarded to John M. Kratz, MD, Membership Committee Chairman, Southern Thoracic Surgical Association, 401 North Michigan Avenue, Chicago, IL 60611-4267.
Hendrick B. Barner, M D Secretary-Treasurer Southern Thoracic Surgical Association 401 North Michigan Avenue Chicago, IL 60611-4267