Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Heterogeneous tissue characterization using ultrasound: a comparison of fractal analysis backscatter models on liver tumors

Heterogeneous tissue characterization using ultrasound: a comparison of fractal analysis... Assessing tumor tissue heterogeneity via ultrasound has recently been sug- gested for predicting early response to treatment. The ultrasound backscat- tering characteristics can assist in better understanding the tumor texture by highlighting local concentration and spatial arrangement of tissue scatter- ers. However, it is challenging to quantify the various tissue heterogeneities ranging from ne-to-coarse of the echo envelope peaks in tumor texture. Lo- cal parametric fractal features extracted via maximum likelihood estimation from ve well-known statistical model families are evaluated for the purpose of ultrasound tissue characterization. The fractal dimension (self-similarity measure) was used to characterize the spatial distribution of scatterers, while the Lacunarity (sparsity measure) was applied to determine scatterer num- ber density. Performance was assessed based on 608 cross-sectional clinical ultrasound radio-frequency images of liver tumors (230 and 378 demonstrat- ing respondent and non-respondent cases, respectively). Cross-validation via leave-one-tumor-out and with di erent k -folds methodologies using a Bayesian classi er were employed for validation. The fractal properties of the backscattered echoes based on the Nakagami model (Nkg) and its extend four-parameter Nakagami-generalized inverse Gaussian (NIG) distribution achieved best results { with nearly similar performance { for characterizing liver tumor tissue. Accuracy, sensitivity and speci city for the Nkg/NIG Corresponding Author: E-mail. o.alkadi@ju.edu.jo. Preprint submitted to Ultrasound in Medicine and Biology December 23, 2019 arXiv:1912.09903v1 [eess.IV] 20 Dec 2019 were: 85.6%/86.3%, 94.0%/96.0%, and 73.0%/71.0%, respectively. Other statistical models, such as the Rician, Rayleigh, and K-distribution were found to not be as e ective in characterizing the subtle changes in tissue texture as an indication of response to treatment. Employing the most rel- evant and practical statistical model could have potential consequences for the design of an early and e ective clinical therapy. Keywords: Ultrasound, tissue characterization, liver tumor, RF envelope, fractal analysis, texture analysis 2 Introduction Ultrasound tissue characterization can provide useful quantitative assess- ments for understanding the state of biological disease [1]. With advancement in medical image analysis, it is becoming a promising non-invasive technique for early detection of tumor response to treatment [2, 3, 4]. It has the ad- vantage of deriving parameters that can represent tissue properties in a fast, non-ionizing, easily operated, and cost-e ective way compared to other con- ventional follow-up imaging techniques. Soft tissue pathologies in the form of lesions tend to have distinct scattering patterns to that of normal tis- sue structure, and the associated acoustic properties could characterize the concentration of scatterers and micro-structures; which is an indication of di erent tissue types. Biological tissue ultrasonic modeling followed by echo signal analysis can facilitate heterogeneity examination of tumor texture. The interaction of an acoustic wave with di erent tissue regions can be modeled by the backscattered radio-frequency (RF) signal. Tissue proper- ties based on the scatterer number density and spatial distribution can be derived subsequently for analysis. There are several approaches for which useful information can be extracted from the RF signal. One approach uses the local power spectral density to estimate the integrated backscatter and attenuation coecient [5, 6, 7, 8], or to measure the mean central frequency and scatterer size [9, 10, 11, 12]. Textural properties of the tissue spatial arrangement can also be estimated from the envelope-detected RF image [13, 14]. As the rst-order statistical properties of the backscattered RF sig- nal rely on the number density and spatial distribution of scatterers [15, 16] { which maybe coherent, random or a mixture of both { it would be dicult to account for all scatterer conditions using the former approaches. Therefore others have investigated the probability density function of the backscattered echoes and proposed to account for the number, size, spacing and regularity of the scatterers in tissue. An overview of the various statistical distributions for modeling the envelope-detected RF signal can be found in [17]. For the latter approach, the main objective is to provide a better characterization of the fundamental elements that form the coarse textural patterns, namely speckles formed from the backscattered echoes. The speckle local arrange- ment represents the various scatterer concentrations and spatial distributions occurring in tissue; ranging from a fully developed, partially developed, to a coherent speckle pattern. In cases when there are many randomly located scatterers per resolution cell (i.e. fully developed speckle), the envelope sig- 3 nal statistics would follow a square root of exponential distribution, known as Rayleigh distribution [15]. The model can be further subdivided into pre-, Rayleigh, and post-Rayleigh for characterizing heterogeneous, homogeneous, and periodic textures, respectively [18, 19, 20]. Non-Rayleigh distributions can be observed when the scatterers become less condensed or having a struc- ture with some regularity. That is, when the number of scatterers in a reso- lution cell is small (i.e. partially developed speckle), the K-distribution was found to be more e ective in modeling the pre-Rayleigh statistics of the RF envelope [21]. The RF envelope statistics in this case would have a shape resembling a square root of the product of a gamma and exponential distribu- tions. Whereas if the scatterers become organized with some periodicity (i.e. coherent speckle), the Rician distribution { which is the generalization of the Rayleigh distribution { is more appropriate for characterizing the underlying regular structures in tissue [22]. Although the prior models account for speci c aspects of scatterer lo- calization, they are not general enough to model the various tissue texture conditions. This is true when a high degree of variability in the scattering cross-section associated with a low number of scatterers exist. In practice it is very common to encounter non-Rayleigh conditions of ultrasound backscat- ter, such as mixtures of di use and coherent or periodically aligned scatterers in tissue micro-structure. Therefore a comprehensive approach is sought as in the generalized K-distribution [23] and homodyne K-distribution [24, 25]. A third parameter, which represents the envelope of the coherent signal, was added to the e ective number of scatterers and energy of the random scatter- ers to account for post-Rayleigh conditions. However a drawback lies in the analytical complexity of these model generalizations, rendering the process of parameter estimation computationally expensive. Other models which also apply a generalization approach to the Rayleigh and Rice distributions to better- t the backscattered echo include: Weibull [26], Rician inverse Gaus- sian [27], and generalized gamma [28]. On the other hand, the Nakagami distribution family can provide a simpler and general model for ultrasonic tissue characterization [29]. The regularity of the scatterer spacing are taken into consideration besides the scatterer density and amplitude, making it pos- sible to account for hypoechoic and hyperechoic structures [30]. The model shape equivalent to a scaled square root of a gamma distribution can better represent the backscattered RF signal envelope, and can be easily ne-tuned via the shape Nakagami parameter to represent low and high number of scat- terer densities with minimal error [31]. The model was further generalized 4 as a Nakagami-generalized inverse Gaussian distribution in [32] by including an additional shape adjustment parameter to account for the tails of the density function. However this generalization was not investigated with real tissue, where scatterers tend to have a high degree of variability in scattering cross-sections. All aforementioned statistical models of the backscattered echo envelope in the literature claim a better characterization of texture anisotropic proper- ties. A large number of articles address the problem of soft tissue characteri- zation and diagnosis from ultrasound images of various internal organs, such as in kidneys [33], liver [34], breast [35], gallbladder [36], pancreas [37], spleen [38] and abdominal aorta [39] all of which demonstrates practical examples, but not limited to, of recent clinical work . However, to the best of our knowledge, the study of the performance of the most well-known statistical models for characterizing ultrasonic liver tumor tissue has not been investi- gated. Local parametric fractal features extracted via maximum likelihood estimation for each statistical model will be used for liver tumor texture clas- si cations. In this paper a fractal approach for evaluating the performance of the di erent statistical models would be sought for an automated detection of liver tumor response to chemotherapy treatment. The approach relates the fractal characteristics of the tissue scatterers to the underlying statis- tical properties derived from the RF envelope-detected signal. The fractal dimension which represents the degree of self-similarity and the derived La- cunarity which indicates the level of spaces within the texture were related to the scatterers spatial distribution and number density, respectively. The fractal features are extracted via a multimodal statistical distribution and results fed to a classi er for automated classi cation. The aim of the paper is to demonstrate the ecacy in modeling the underlying tumor physiologi- cal changes from ultrasound parametric images of scatterer tissue properties. Tumor tissue scatterer characterization is a challenging task due to the dis- ease chaotic behavior. Therefore we will address the issue of best model selection, and whether analytically complex models are really necessary for better characterizing complex tissue, such as liver tumors. The paper is organized as follows: the preference of working with RF envelope-detected signal rather than B-mode images is brie y explained in the ultrasound data representation section, and followed by modeling the backscattered RF signal to cover the di erent scatterer conditions and densi- ties. Then the multiscale feature extraction section discusses characterizing the fractal patterns of the tissue scatterers for automated classi cation. The 5 results and discussion sections list and discuss the major ndings, and the paper is summarized in the conclusion section. Ultrasound data representation RF parametric image acquisition The acquired ultrasound beams run through several signal and image pre-processing steps before they can be presented as a radio-frequency (RF) matrix, and subsequently in a gray scale B-mode image, see Fig.1.The ul- trasound operator may also choose to further optimize the visibility by in- creasing the gain of the re ected signals with increasing time from the trans- mitted pulse (i.e. time-gain compensation). While the B-mode scans could be suitable for qualitative analysis in describing the tissue structure, some aspects could be in fact obscured or modi ed, and would not realistically represent the condition in the raw data. The post-processing functions that have been applied manipulate the image on the screen, without improving the signal quality itself, or the fundamental signal to noise ratio. The RF envelope-detected signal can better provide quantitative analysis and bal- ance the tradeo between preserving most of the ultrasound data unaltered and with minimal ltering (i.e. only removal of high frequency oscillations). Many works in the literature has shown the usefulness of RF data for clinical analysis as mentioned in Section 1, therefore this work utilizes the envelope- detected RF data for ultrasound tissue characterization of liver tumors. Clinical liver tumor cross-sectional images Cross-sectional images of liver tumors undergoing chemotherapy treat- ment obtained as part of an ethically approved prospective study was used to validate the clinical signi cance of the di erent statistical models. A to- tal of 608 cross-sectional images (230 from tumors responded to treatment categorized as respondent acquired from 11 patients, and 378 from tumors progressed categorized as non-respondent acquired from 22 patients) were obtained using a diagnostic ultrasound system (z.one, Zonare Medical Sys- tems, Mountain View, CA, USA) with a 4 MHz curvilinear transducer and 11 MHz sampling. The target tumor was initially imaged using the optimal acoustic window as determined by an experienced operator with 5 years experience of liver ultrasound imaging. The depth of imaging was kept constant for all study acquisitions. Other imaging settings including scanning frequency, gain and 6 Figure 1: Illustration of ultrasound RF to B-mode data conversion process. time gain compensation were also kept constant. The operator then posi- tioned the imaging transducer to ensure the target tumor is optimally placed at the center of the image. While maintaining stable skin contact position, the transducer was angled to image the most cranial aspect of the tumor. The transducer was then swept in a fan motion towards the caudal aspect of the tumor, capturing multiple cross sectional images through the target tumor. The procedure was repeated at all subsequent time points during the moni- toring period. Since the imaging protocol was constant, the tumor should be at identical depth every time it is imaged throughout the monitoring period. Each dataset was acquired prior to commencement of chemotherapy. Re- sponse to treatment was determined based on conventional computed tomog- raphy follow up imaging as part of the patient standard clinical care based on the response evaluation criteria in solid tumors (RECIST) [40]. The base- line cross-sectional imaging was compared against those performed at the end of treatment according to the RECIST criteria to determine response to treatment for each target tumor. A tumor was classi ed as responsive if 7 categorized as partial response and non-responsive if no change or disease demonstrated progression. Cross-validation was performed based on a leave- one-tumor-out approach for a total of 40 volumetric (stacks of segmented 2D slices) tumors, for which 13 were responsive and 27 did not respond to chemotherapy treatment. Imaging was done by two radiologists to provide gold standard and training data. The acquisition of the dataset in this work did not in uence the diagnostic process or the patients' treatment, and in- formed consent was received from each participant in the study. Modeling the backscattered RF signal Ultrasound backscattering in tissue can be considered as a random pro- cess. Many statistical models have been applied for describing the proba- bility density function of the envelope-detected RF signal. Evaluation has been made by ve major statistical distributions that are widely used in the literature to describe the various aspects of tissue characteristics. Rayleigh distribution The Rayleigh distribution ideally suits the condition when the ampli- tudes of the individual backscattered signals are assumed to be randomly distributed, i.e. speckle pattern is \fully developed". The n -dimensional Rayleigh distribution [17] can be seen as a continuous probability distribu- tion for positive-valued random variables de ned n=2 2 1 2 2 2 n1 x =2 P xj = x e ; (1) ray (n=2) 2 where x represents the amplitude of the signal,  > 0 is the scale parameter of the distribution, and denotes the Euler gamma function. The case when n = 2 corresponds to the Rayleigh distribution. Rician distribution The Rician or Rice distribution model combines a coherent component to the backscattered signal that underlies the Rayleigh distribution. It has the capability of modeling unresolved structures exhibiting a periodic pattern in the spatial scatterer distribution [41]. An n -dimensional Rician distribution [22] can be expressed as n=2 x  2 2 2 +x =2 2 ( ) P xj;  =  I x e ; (2) ric n=21 2 2 8 where x represents the amplitude of the signal, the scale parameter  > 0 and non-centrality parameter   0 are real numbers, n is the dimension and I denote the modi ed Bessel function of the rst kind of order p. When n = 2 the distribution would correspond to a Rician distribution.The special case where  ! 0 would yield the Rayleigh distribution. K-distribution The K-distribution can be considered as a generalization of the Rayleigh distribution that can account for non-uniformities in the spatial scatterer distribution. This property allows the modeling of clusters of scatterers or when the e ective number of scatterers per resolution cell is small. The K-distribution [17] is de ned by 1+n=2 4x 2 P xj ; = K x ; (3) kd n=2 ( +n=2)=2 2 (2 ) ( ) (n=2) where > 0,  > 0 are the shape and scale parameters, respectively, and K denotes the modi ed Bessel function of the second kind of order p. The K-distribution includes a Rayleigh distribution as a special case for = 1. Nakagami distribution This is another family of distributions that can model the ultrasonic backscattered envelope for various scattering conditions and scatterer densi- ties. The Nakagami density function [42] is de ned as 2m 2m1 mx = P (xjm; ) = x e ; (4) nkg (m) for x  0, where is the Euler gamma function. The real numbers m > 0 (related to the local backscattered energy) and > 0 (related to the local concentration of scatterers) are called the shape and scaling parameters, respectively. Similarly to the Rayleigh distribution, the envelope of the RF signal x follows a gamma distribution. By ne-tuning the shape of the distribution parameter m, other statistical distributions can be modeled, such as, an approximation of the Rician distribution (i.e. post-Rayleigh) for m > 1, a Rayleigh distribution for the special case when m = 1, and when m < 1 a K-distribution (i.e. post-Rayleigh) for modeling the backscattered echoes with less dense scatterers. Thus the SNR of the Nakagami distribution can take any positive value. 9 Nakagami-generalized inverse Gaussian distribution Based on a four parameter model, the Nakagami-generalized inverse of Gaussian distribution (NIG) distribution [32, 17] is an extended form of the Nakagami distribution to better cover the coherent scattering conditions. It is de ned by 2 (m= 2m1 P (xjm; ; ; ) = x NIG p =2 K  (m) (m)=2 2mx K  + ; (5) 2mx + where m, ,  and  are positive real numbers with the condition  > 0; 0, and K is modi ed Bessel function of the second kind. By having the four parameter values of the NIG distribution model set to speci c conditions, various well-known distributions can be recovered such as, Rayleigh (m = 1;  > 0; ! 1), Nakagami (m > 1;  > 0; ! 1), K-distribution (m = 1;  > 0; ! 0), and Nakagami-gamma (m > 0;  > 0; ! 0). Besides the NIG distribution, other complex models such as the Generalized- K, the Homodyned-K, and the Rician inverse of Gaussian distributions exist in the literature for ultrasound tissue characterization. However they were not used as the simpler and more general models of Nakagami family { in its 2 and 4-parameter extended form { which under di erent limiting condi- tions can approximate the known distributions, and hence can cope with the various scattering conditions related to the coherent-to-di use signal power ratio and scatterer clustering, see illustration in Fig.2. Multiscale feature extraction Parametric image generation Based on a voxel-by-voxel approach, the statistical model parameters are estimated from the RF envelope data. Each estimated parameter in the generated parametric image corresponds to a voxel { short segment { of the RF signal. As we are dealing with a tumor volume, the RF enve- lope signal was presented as a matrix, and several matrices represent the tumor volume. Then the di erent statistical distribution models are t- ted and then their associated model parameters estimated by maximum likelihood estimation. Namely, the maximum likelihood estimate  (v) for 10 Scatterer characterization Figure 2: An illustration representing the coverage of the scatterer distribution vs scat- terer number density by the 5 statistical distributions employed in this work for tissue characterization. a density function f (v ; : : : ; v =) when  is a vector of parameters for 1 m a certain distribution family , estimates the most probable parameters (v) = argmax D (=v ; : : : ; v ), where D (=v) = f (v=) ;  2  is the 1 m score function. Thereby, a number of parametric images l , where r = 1; 2; or 4 for a mono, bi, and quad-parametric model, respectively, are generated for each corresponding RF envelope image of the applied distribution model. Herein, l refers to the statistical model parameters estimated locally from the neighborhood of the RF envelope data by a square{sliding window with its size optimized as in [43] and centered on each point in the RF data matrix. Also the same RF envelope data was used to create the di erent parametric images for each statistical model. That is, creating for each model maps of the distribution parameters throughout the entire scanned region (i.e. each parameter map being an image). An interesting point to consider is the di erence between the spatial dis- tribution of tissue scatterers, and the structure of the texture of the para- metric ultrasound images. The former results from the interaction of the 11 sound energy with the liver tumor tissue leading to scatter in many direc- tions. Complex structure of tissue scatterers and their small size as com- pared to the ultrasound wavelength, gives the speckle patterns a Rayleigh scattering behavior. The spatial distribution of these tissue scatterers is not related to any particular ultrasound imaging system, but is a property of the tissue structure and the ultrasound frequency used which dictates the scat- tering length scale. On the other hand, texture of the parametric ultrasound images, which shows di erent spatial variations of speckle intensity, is depen- dent on the used ultrasound imaging system. Usually a set of proprietary lters { speci c to the ultrasound device manufacturer { are used for post- processing operations, which may result in loss of information. Therefore the RF envelope-detected data in this work did not undergo any post-processing ltering or time-gain compensation operations. Fractal features Various statistical signatures can be derived from the acoustic properties of the backscattered RF and envelope-detected signals. An e ective way to investigate the local arrangement of the scatterer concentrations and spatial distributions occurring in tissue is using a multifractal analysis approach. The fractal dimension (FD) can indicate the degree of \self-similarity" in the spatial distribution of tissue scatterers, and has shown many promising results in tissue characterization from di erent imaging modalities [44, 45, 43, 46, 47]. Also the Lacunarity, which is a measure derived from the texture of the FD, re ects the gaps within the tissue texture, and hence measures the number density of tissue scatterers. Both signatures, the FD de ned in (6) { estimated via the fractal Brownian motion which is a non-stationary model known for its capability for describing random phenomena [48] { and its counterpart Lacunarity which is the standard deviation divided by the mean of a fractal image F using a sliding box-counting algorithm as shown in (7), are considered useful for revealing the mixtures of spatial homogeneity and heterogeneity properties of tissue texture in the generated parametric images. log (N ) FD = ; (6) log (1=s) P P M1 N1 2 1=MN F (m; n) m=0 n=0 L = 1 (7) P P M1 N1 1=MN F (k; l) k=0 l=0 12 where N is the number of self-similar shapes and s is the corresponding scaling factor; M and N in (7) are the size of the fractal parametric image F . In this work, the FD and derived Lacunarity signatures are estimated in a multiscale fashion after applying a Daubechies wavelet packet analysis de- composition technique, where the texture variations of the subbands at each level of decomposition serve as the extracted fractal feature pattern. The Daubechies wavelet can account for self-similarity and signal discontinuities, making it useful for characterizing signals exhibiting fractal patterns [49]. An orthogonal 8-tap Daubechies lter was used to obtain the wavelet packets by expanding the basis having the most signi cant fractal signature rather than energy [50]. The fractal features at the di erent resolution scales give more information on the local concentration and spatial arrangement of the tissue scatterers { as higher FD signature values indicate a more heterogeneous texture (i.e. distribution of scatterers are random), and vice versa [51, 52]; whereas Lacunarity signature values operate oppositely, giving a low value when the scatterers are large, viz. less gaps between scatterers per resolution cell, see Fig.3. An illustrative example of liver tumor tissue characterizing by the di erent statistical models is shown in Fig.4. The rst column shows the distribution and number density of the tissue scatterers modeled by the di er- ent statistical distributions, while the second column represents the textural patterns of the modeled tissue scatterers in a multifractal analysis approach. l l At the end of the feature extraction stage, a feature vector = = (w ; w ; 1;1 1;2 : : : ; w ) would consist of all selected fractal signatures w for a speci c model r;q parameter r and wavelet subband q from each RF envelope image l. A data pipeline analysis of the fractal feature extraction process is illustrated in Fig.5. Pattern classi cation In order to assess the quality of the extracted features from the di erent statistical distribution models, a na ve Bayesian classi er (NBC) was used for classi cation. The NBC is a robust learning method which can work well even when the data is non-normally distributed [53], and the assumption of attribute independence allows the NBC classi er to estimate the classi cation accuracy using less training data as compared to other classi ers. This fast and simple classi er was reported to perform well in practice even with the presence of strong attribute dependence [54]. According to Bayesian theory, 13 Fractal characterization Figure 3: Tissue characterization using the fractal properties of the scatterers distribution and clustering: The fractal dimension measures the level of self-similarity (i.e. spatial distribution), while the Lacunarity assesses the degree of heterogeneity (i.e. denseness) in tissue texture. each fractal feature = ; f = 1; : : : ; n, would be assigned to class i, if P (C j=) f i is maximized over all classes P (C j=) > P (C j=)8j 6= i. i j To nd P (C j=), NBC can be represented as: P (C ) P (= jC ) i f i f =1 P (C j=) = (8) P Q k n P (C ) P (= jC ) i f i i=1 f =1 where P (C j=) is the a posteriori probability of assigning class i given feature th vector =, P (= jC ) is the probability density function of = within the i f i class C for a total number of k classes, P (C ) and P (=) are the a priori i i probability of class C and feature vector =; respectively. Results The extracted fractal features from the texture of the various parametric images { which where derived from the RF envelope-detected signal of the dif- ferent distribution models {are compared in Tables 1, 2, and 3, where values 14 Figure 4: Characterizing a segment of a liver tumor tissue: (from left to right) model para- metric image and corresponding fractal image for (a) Nakagami, (b) Nakagami-generalized inverse Gaussian, (c) Rayleigh, (d) Rician, and (e) K-distribution statistical models, re- spectively. 15 Figure 5: Data pipeline analysis explaining the fractal feature extraction process. 16 in bold indicate the best achieved performance. A detailed classi cation per- formance of the 3D clinical ultrasound liver tumor test set via a leave-one-out cross-validation approach is shown in Table 1. The tissue characterization based on the Nakagami (Nkg) distribution and its 4-parameter extended ver- sion (NIG distribution) outperformed the other statistical models in terms of classi cation performance. Both Nkg and NIG had nearly equal classi ca- tion accuracy and the area under the receiver operating characteristics (ROC) curve with 85.6% & 86.3%, and 83.2% & 83.5%, respectively. Although NIG had the highest sensitivity (true-positive rate), the Nkg achieved relatively better in the statistical measures of speci city (true-negative rate), precision, and false-positive (FP) rate. The classi cation performance of the Rayleigh distribution was the third among all ve distributions; nevertheless achiev- ing equal statistical measures regarding speci city and FP rate with the NIG model. The Dice similarity coecient arms the accuracy results and shows how similar the predicted and observed responses are in regard to chemother- apy treatment. Although the K-distribution was the least sensitive in terms of classi cation performance { next to Rician distribution, it recorded the best speci city score as compared to the other distributions. In order to avoid over tting and to give some insights on how would the di erent distribution models based on the scatterers fractal characteristics generalize, the extracted features were further validated. Table 2 and 3 show the detailed classi er performance on the test set for 10- and 20-fold cross- validation (results are the mean  standard deviation of the performance over 60 runs), respectively. Table 1: Detailed Classi cation Performance (leave-one-out cross-validation) for di erent statistical models of RF ultrasound for liver tissue characterization Classi cation RF-based stochastic distribution Performance Nkg Ric Ray NIG Kd FP rate 0.271 0.327 0.289 0.289 0.336 Sensitivity 0.940 0.890 0.920 0.960 0.880 Speci city 0.730 0.670 0.710 0.710 0.770 Accuracy 0.856 0.810 0.838 0.863 0.798 Precision 0.850 0.810 0.830 0.840 0.795 Dice SC 0.922 0.892 0.912 0.926 0.888 ROC Area 0.832 0.781 0.814 0.835 0.773 17 Table 2: Detailed Classi cation Performance (10-fold cross-validation) for di erent statistical models of RF ultrasound for liver tissue characterization Classi cation RF-based stochastic distribution Performance Nkg Ric Ray NIG Kd FP rate 0.280  0.076 0.345  0.113 0.285  0.089 0.332  0.045 0.380  0.128 Sensitivity 0.924  0.720 0.887  0.655 0.911  0.715 0.955  0.668 0.870  0.619 Speci city 0.720  0.924 0.655  0.887 0.715  0.911 0.668  0.955 0.620  0.872 Accuracy 0.845  0.007 0.797  0.009 0.836  0.007 0.845  0.009 0.780  0.011 Precision 0.840  0.856 0.803  0.785 0.836  0.836 0.821  0.905 0.790  0.754 Dice SC 0.916  0.004 0.887  0.005 0.910  0.004 0.916  0.006 0.870  0.007 ROC Area 0.822  0.009 0.771  0.010 0.813  0.008 0.812  0.011 0.75  0.012 Table 3: Detailed Classi cation Performance (20-fold cross-validation) for di erent statistical models of RF ultrasound for liver tissue characterization Classi cation RF-based stochastic distribution Performance Nkg Ric Ray NIG Kd FP rate 0.274  0.073 0.332  0.112 0.283  0.084 0.308  0.041 0.361  0.124 Sensitivity 0.927  0.726 0.888  0.668 0.916  0.717 0.959  0.692 0.876  0.639 Speci city 0.726  0.927 0.668  0.888 0.717  0.916 0.692  0.959 0.639  0.876 Accuracy 0.849  0.007 0.803  0.005 0.839  0.007 0.856  0.008 0.785  0.008 Precision 0.843  0.862 0.809  0.789 0.837  0.843 0.832  0.915 0.794  0.765 Dice SC 0.918  0.004 0.891  0.003 0.912  0.004 0.922  0.004 0.879  0.005 ROC Area 0.826  0.007 0.778  0.006 0.816  0.008 0.826  0.009 0.758  0.010 Since we are concerned more with the ability of an early chemotherapy treatment to identify initial tissue response to treatment, namely the sensi- tivity of the treatment procedure, Fig.6 shows the di erence in the sensitivity measure between the raw RF parameters and the fractal-based measures ex- tracted from the parametric images of the RF envelope detected signal. The Wilcoxon signed-rank test on paired overall sensitivity for the fractal and parametric-based of the statistical models shows there is signi cant di er- ences between the two approaches (p < 0:05). For completeness, the models accuracy are shown in Fig.7 as well. 18 Sensitivity Comparison Figure 6: Sensitivity comparison of the di erent statistical models for tissue character- ization based on the using the estimated model parameters (parametric-based) and by the fractal properties of the tissue scatterers(fractal-based). The models sensitivity was cross-validated via (a) leave-one-out, (b) 10-fold, and (c) 20-fold. Discussion Heterogeneity in the tumor tissue scatterers could span di erent scatter- ing conditions. Regions within the tumor tissue which respond to treatment 19 Accuracy Comparison Figure 7: Cross-validation accuracy comparison of the di erent statistical models for tissue characterization based on the using the estimated model parameters (parametric-based) and by the fractal properties of the tissue scatterers(fractal-based). The models accuracy was cross-validated via (a) leave-one-out, (b) 10-fold, and (c) 20-fold. might exhibit di erent statistical properties to that of the non-respondent counterpart. Thus it is essential to evaluate the discriminative abilities of 20 the statistical models while simultaneously taking into account all possible scatterer conditions and densities. The sum of the individual backscattered signals from the various scatter- ing regions within the biological tissue can be used for RF signal modeling. In the case of a fully developed speckle pattern, the size of the scatterers tend to be small in comparison with the wavelength and also have a disor- ganized structure, i.e. the spatial distribution of scatterers is random [55]. This can be considered as a typical case for tumor tissue, where the tumor angiogenesis process causes for the newly induced blood vessels to grow in a chaotic manner, and therefore tissue scatterers to have a random nature. The relation between angiogenesis and randomness of tissue scatterers can be attributed to the chaotic way that tumors build their network of new blood vessels, and thus giving tumorous tissue a \rougher" appearance than non-tumorous. These angiogenesis networks tend to be leaky and disorga- nized, unlike blood vessels in normal tissue. The chaotic behavior introduces a degree of randomness in appearance or \roughness" and gives a higher fractal dimension value compared to necrotic tissue. However spatial scat- terer distribution in soft tissue is not strictly random and may not originate from identical scatterers, also it is common to encounter di erent degree of variability in tumor tissue as scatterers vary in shape and size, with a mix- ture of partially developed and coherent speckle patterns. This compound situation becomes more distinct once certain regions within the tumor start to respond to treatment, which re ects in the statistics of the RF backscat- tered signal. Moreover, after generating the parametric images from the RF envelope echo, the fractal characteristics of tissue scatterers can provide addi- tional information on the local spatial distribution and number density. Fig.6 and Fig.7 illustrate an improved performance when a fractal characterization is applied to the parametric images. The fractal dimension is analogous to the coherent-di use ratio in Fig.2, where it tends to improve distribution separability from a texture perspective by assessing the scatterers random- periodic properties. Also the Lacunarity of the textures in the parametric images further re nes the scatterer densities and measures the degree of het- erogeneity in tumor tissue. The fractal approach shows a better sensitivity for all distribution models in Fig.6 with the best improvement achieved by the Nakagami model, while the NIG model recorded the highest sensitiv- ity. Similarly in Fig.7, where the accuracy is remarkably improved for the fractal-based images of the Nakagami model, and achieving a nearly compa- rable performance to that of the complex NIG model. A point to note is the 21 decrease in the K-distribution accuracy for the cross-validations in Fig.7(b) and (c). This indicates that the fractal approach is incapable of extracting useful information from the tissue scatterers characterized by the model, and hence giving inaccurate characterization of tumor heterogeneity. In some typical conditions of ultrasound tissue characterization, the backscat- tered signal cannot be assumed to be Rayleigh distributed, such as when the number of scatterers per resolution cell are not large enough or when there is a lack of randomness in the scatterers localization (e.g. due to period- icity, structure, or clustering) [41]. The former conditions suggest that the Rayleigh distribution may not approximate e ectively the statistical charac- teristics of the echo envelope when tumor tissue regions becomes less dense and having a more consistent structure. Although tumor tissue is known to have a chaotic nature [56], the change in the tissue scatterer textural proper- ties in a sense of becoming less random (e.g. inducing much less angiogenesis), or even the possible reduction in the number of e ective scatterers in some parts of the tumor tissue (e.g. more necrotic regions) could be attributed to initial response to treatment and tumor becoming less aggressive. This explains why the Rayleigh distribution in Table 1 did not perform as well as the Nakagami distribution family which can take into account such variabil- ity in both tissue scatterer distribution and number density. On the other hand, it is not expected to encounter periodic patterns within the tumor tissue due to its chaotic nature, therefore dominance of periodic or coher- ent patterns in the modeled tissue texture is very rare, which explains the reduced classi cation performance based on the Rician distribution. As a result, the spatial scatterer distribution non-uniformities, such as clusterings in scatterers or regions with smaller e ective numbers of scatterers cannot be e ectively modeled. Also the K-distribution can be used to model the spacing or regularity in the spatial scatterer distribution, such as in partially developed speckle based on a low-average number of e ective scatterers per resolution cell; however, it had the least classi cation performance in terms of predicting liver tumor response to chemotherapy treatment. This can be seen as tissue scatterers in liver tumors tend to have a high degree of num- ber density which the model does not cover, and relying on characterizing the tissue randomness only might not suce. Also the results in Table 1, 2, and 3 show the NIG recorded nearly similar performance in terms of clas- si cation accuracy to that of the Nakagami model. The main advantage of the extended NIG model over the bi-parameter Nakagami model is the bet- ter coverage of the coherent-di use region for low scatterer number density. 22 Since there is not much obvious structural or periodic components in the tissue echo signal of the liver tumor, both models exhibited similar behavior. Moreover, the higher sensitivity values for the NIG model suggest that it is better in predicting response to chemotherapy treatment (i.e. aggressive tu- mors that are responding to treatment), while the Nakagami model is more precise and has the least overall FP rate. Finally, the relatively low digital sampling rate could be a potential lim- itation. The e ect of sparse sampling of the analog signal on the fractal analysis of the RF signal needs to be further investigated. Also as tumors may be localized at varying depths, this may decrease the amplitude of the RF data. Therefore the performance of the derived model parameters based on the fractal analysis technique as a function of depth could serve as future work. Furthermore, the study of the link between the FD derived from para- metric images of weak scattering sources to the spatial distribution of tissue scatterers would assist in better understanding the robustness of the fractal analysis approach. Conclusions This paper presents an evaluation of di erent statistical models that cover di erent aspects of the tissue scatterers distributions and densities. The aim is to understand which model can give a better- t of the complex nature of liver tumors by providing discriminative features for supporting clinical diagnosis as part of healthcare delivery. The tumor heterogeneity was as- sessed via its tissue fractal properties derived from the RF envelope-detected signal and used as features for assessing classi cation performance. The Nak- agami distribution family was the most e ective in characterizing the liver tumor tissue, with the simple bi-parametric model achieving equal accuracy and better precision as compared with the complex 4-parameter extended version. Moreover, fractal analysis of the echo signal was found to give ad- ditional information about the underlying tissue structure, and hence better indicates response to chemotherapy treatment. These ndings suggest that the simple and general models of the Nakagami distribution family along with the fractal analysis of the RF parametric images have promising potential in providing improved ultrasound tissue characterization of liver tumors. 23 Acknowledgements This work was support by the Engineering and Physical Sciences Research Council and Wellcome Trust Grant WT 088877/z/09/z. The authors would like to thank the anonymous reviewers for their constructive comments and suggestions to improve the quality of the paper. 24 References [1] J. Mamou, M. L. Oelze, Quantitative Ultrasound in Soft Tissues, Springer, 2013. [2] G. Czarnota, A. Sadeghi-Naini, N. Papanicolau, O. Falou, R. Dent, S. Verma, M. Trudeau, J.-F. Boileau, J. Spayne, S. Iradji, E. Sofroni, J. Lee, S. Lemon-Wong, M. Ya e, M. Kolios, Quantitative ultrasound evaluation of tumor cell death response in locally advanced breast cancer patients to chemotherapy treatment administration, The Journal of the Acoustical Society of America 133 (5) (2013) 3539. [3] A. Sadeghi-Naini, O. Falou, G. J. Czarnota, Characterizing tumor het- erogeneous response to chemotherapy using low-frequency ultrasonic spectroscopy, The Journal of the Acoustical Society of America 133 (5) (2013) 3584. [4] A. Sadeghi-Naini, N. Papanicolau, O. Falou, J. Zubovits, R. Dent, S. Verma, M. Trudeau, J. F. Boileau, J. Spayne, S. Iradji, E. Sofroni, J. Lee, S. Lemon-Wong, M. Ya e, M. C. Kolios, G. J. Czarnota, Quan- titative ultrasound evaluation of tumor cell death response in locally advanced breast cancer patients receiving chemotherapy, Clinical Can- cer Research 19 (8) (2013) 2163{2174. [5] L. Sannachi, H. Tadayyon, A. Sadeghi-Naini, W. Tran, S. Gandhi, F. Wright, M. Oelze, G. Czarnota, Non-invasive evaluation of breast can- cer response to chemotherapy using quantitative ultrasonic backscatter parameters, Medical Image Analysis 20 (1) (2015) 224{236. [6] N. Rubert, T. Varghese, Scatterer number density considerations in ref- erence phantom-based attenuation estimation, Ultrasound in Medicine and Biology 40 (7) (2014) 1680{1696. [7] K. Nam, J. A. Zagzebski, T. J. Hall, Simultaneous backscatter and attenuation estimation using a least squares method with constraints, Ultrasound in Medicine and Biology 37 (12) (2011) 2096{2104. [8] B. Banihashemi, R. Vlad, B. Debejevic, A. Giles, M. C. Kolios, G. J. Czarnota, Ultrasound imaging of apoptosis in tumor response: Novel preclinical monitoring of photodynamic therapy e ects, Cancer Research 68 (20) (2008) 8590{8596. 25 [9] S. L. Bridal, P. Fornes, P. Bruneval, G. Berger, Parametric (integrated backscatter and attenuation) images constructed using backscattered radio frequency signals (25-56 MHz) from human aortae in vitro, Ultra- sound in Medicine and Biology 23 (2) (1997) 215{229. [10] E. P. Nordberg, T. J. Hall, E ective scatterer diameter estimates for broad scatterer size distributions, Ultrasonic imaging 37 (1) (2015) 3{ [11] R. Lavarello, M. Oelze, Quantitative ultrasound estimates from popula- tions of scatterers with continuous size distributions: E ects of the size estimator algorithm, IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control 59 (9) (2012) 2066{2076. [12] R. K. Saha, M. C. Kolios, E ects of cell spatial organization and size distribution on ultrasound backscattering, IEEE Transactions on Ultra- sonics Ferroelectrics and Frequency Control 58 (10) (2011) 2118{2131. [13] N. Bouhlel, S. Sevestre-Ghalila, Nakagami markov random eld as tex- ture model for ultrasound rf envelope image, Computers in Biology and Medicine 39 (6) (2009) 535{544. [14] T. Klein, M. Hansson, N. Navab, Spatial statistics based feature descrip- tor for rf ultrasound data, 2011 8th IEEE International Symposium on Biomedical Imaging: From Nano to Macro (2011) 33{36. [15] R. F. Wagner, S. W. Smith, J. M. Sandrik, H. Lopez, Statistics of speckle in ultrasound b-scans, IEEE Transactions on Sonics and Ultrasonics 30 (3) (1983) 156{163. [16] M. Pereyra, H. Batatia, Modeling ultrasound echoes in skin tissues us- ing symmetric alpha-stable processes, IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control 59 (1) (2012) 60{72. [17] F. Destrempes, G. Cloutier, A critical review and uniformized represen- tation of statistical distributions modeling the ultrasound echo envelope, Ultrasound in Medicine and Biology 36 (7) (2010) 1037{1051. [18] R. M. Cramblitt, K. J. Parker, Generation of non-rayleigh speckle dis- tributions using marked regularity models, IEEE Transactions on Ultra- sonics Ferroelectrics and Frequency Control 46 (4) (1999) 867{874. 26 [19] P. M. Shankar, R. Molthen, V. M. Narayanan, J. M. Reid, V. Genis, F. Forsberg, C. W. Piccoli, A. E. Lindenmayer, B. B. Goldberg, Studies on the use of non-rayleigh statistics for ultrasonic tissue characterization, Ultrasound in Medicine and Biology 22 (7) (1996) 873{882. [20] R. C. Molthen, P. M. Shankar, J. M. Reid, Characterization of ultra- sonic b-scans using non-rayleigh statistics, Ultrasound in Medicine and Biology 21 (2) (1995) 161{170. [21] P. M. Shankar, V. A. Dumane, J. M. Reid, V. Genis, F. Forsberg, C. W. Piccoli, B. B. Goldberg, Use of the k-distribution for classi cation of breast masses, Ultrasound in Medicine and Biology 26 (9) (2000) 1503{ [22] J. Sijbers, A. J. den Dekker, P. Scheunders, D. Van Dyck, Maximum- likelihood estimation of rician distribution parameters, IEEE Transac- tions on Medical Imaging 17 (3) (1998) 357{361. [23] E. Jakeman, R. J. A. Tough, Generalized k distribution: a statistical model for weak scattering, J. Opt. Soc. Am. A 4 (9) (1987) 1764{1772. [24] V. Dutt, J. F. Greenleaf, Ultrasound echo envelope analysis using a ho- modyned k-distribution signal model, Ultrasonic Imaging 16 (4) (1994) 265{287. [25] J. Mamou, A. Coron, M. L. Oelze, E. Saegusa-Beecroft, M. Hata, P. Lee, J. Machi, E. Yanagihara, P. Laugier, E. J. Feleppa, Three-dimensional high-frequency backscatter and envelope quanti cation of cancerous hu- man lymph nodes, Ultrasound in Medicine and Biology 37 (3) (2011) 345{357. [26] B. I. Raju, M. A. Srinivasan, Statistics of envelope of high-frequency ultrasonic backscatter from human skin in vivo, IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control 49 (7) (2002) 871{882. [27] T. Eltoft, The rician inverse gaussian distribution: A new model for non-rayleigh signal amplitude statistics, IEEE Transactions on Image Processing 14 (11) (2005) 1722{1735. 27 [28] A. S. Tunis, G. J. Czarnota, A. Giles, M. D. Sherar, J. W. Hunt, M. C. Kolios, Monitoring structural changes in cells with high-frequency ul- trasound signal statistics, Ultrasound in Medicine and Biology 31 (8) (2005) 1041{1049. [29] P. M. Shankar, A general statistical model for ultrasonic backscatter- ing from tissues, IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control 47 (3) (2000) 727{736. [30] P. H. Tsui, C. K. Yeh, Y. Y. Liao, C. C. Chang, W. H. Kuo, K. J. Chang, C. N. Chen, Ultrasonic nakagami imaging: a strategy to visualize the scatterer properties of benign and malignant breast tumors, Ultrasound in Medicine and Biology 36 (2) (2010) 209{217. [31] P. H. Tsui, H. Y. Ma, Z. H. Zhou, M. C. Ho, Y. H. Lee, Window- modulated compounding nakagami imaging for ultrasound tissue char- acterization, Ultrasonics 54 (6) (2014) 1448{1459. [32] Karmeshu, R. Agrawal, Study of ultrasonic echo envelope based on nakagami-inverse gaussian distribution, Ultrasound in Medicine and Bi- ology 32 (3) (2006) 371{376. [33] H. P. Wu, N. G. Rognin, T. M. Krupka, L. Solorio, H. Yoshiara, G. Guenette, C. Sanders, N. Kamiyama, A. A. Exner, Acoustic charac- terization and pharmacokinetic analyses of new nanobubble ultrasound contrast agents, Ultrasound in Medicine and Biology 39 (11) (2013) 2137{2146. [34] G. Ghoshal, R. J. Lavarello, J. P. Kemmerer, R. J. Miller, M. L. Oelze, Ex vivo study of quantitative ultrasound parameters in fatty rabbit liv- ers, Ultrasound in Medicine and Biology 38 (12) (2012) 2238{2248. [35] H. Tadayyon, A. Sadeghi-Naini, L. Wirtzfeld, F. C. Wright, G. Czarnota, Quantitative ultrasound characterization of locally advanced breast can- cer by estimation of its scatterer properties, Medical Physics 41 (1). [36] R. E. Kumon, M. J. Pollack, A. L. Faulx, K. Olowe, F. T. Farooq, V. K. Chen, Y. Zhou, R. C. K. Wong, G. A. Isenberg, M. V. Sivak, A. Chak, C. X. Deng, In vivo characterization of pancreatic and lymph node tissue by using eus spectrum analysis: a validation study, Gastrointestinal Endoscopy 71 (1) (2010) 53{63. 28 [37] G. Atiee, H. Kvitko-White, K. Spaulding, M. Johnson, Ultrasonographic appearance of histoplasmosis identi ed in the spleen in 15 cats, Veteri- nary Radiology and Ultrasound 55 (3) (2014) 310{314. [38] B. Roosens, G. Bala, S. Droogmans, J. Hostens, J. Somja, E. Delvenne, J. Schiettecatte, P. Delvenne, V. Caveliers, T. Lahoutte, G. Van Camp, B. Cosyns, Echocardiographic integrated backscatter for assessing re- duction of aortic valve calci cations by r-568 in a rat model of chronic kidney disease, Ultrasound in Medicine and Biology 39 (11) (2013) 2075{ [39] P.-H. Tsui, C.-K. Yeh, C.-C. Chang, Noise e ect on the performance of nakagami image in ultrasound tissue characterization, Journal of Medi- cal and Biological Engineering 28 (4) (2008) 197{202. [40] E. A. Eisenhauer, P. Therasse, J. Bogaerts, L. H. Schwartz, D. Sargent, R. Ford, J. Dancey, S. Arbuck, S. Gwyther, M. Mooney, L. Rubinstein, L. Shankar, L. Dodd, R. Kaplan, D. Lacombe, J. Verweij, New response evaluation criteria in solid tumours: Revised recist guideline (version 1.1), European Journal of Cancer 45 (2) (2009) 228{247. [41] U. Scheipers, Ultrasonic tissue characterization, in: O. Ukimura, I. Gill (Eds.), Contemporary Interventional Ultrasonography in Urol- ogy, Springer London, 2009, pp. 143{177. [42] M. Nakagami, The m-distribution { A general formula of intensity dis- tribution of rapid fading, in: W. C. Ho mann (Ed.), Statistical Methods in Radio Wave Propagation, Pergamon, Elmsford, NY, 1960, pp. 3{36. [43] O. S. Al-Kadi, D. Y. Chung, R. C. Carlisle, C. C. Coussios, J. A. Noble, Quanti cation of ultrasonic texture intra-heterogeneity via volumetric stochastic modeling for tissue characterization, Medical Image Analysis 21 (1) (2015) 59{71. [44] O. S. Al-Kadi, D. Watson, Texture analysis of aggressive and nonag- gressive lung tumor CE CT images, IEEE Transactions on Biomedical Engineering 55 (7) (2008) 1822{1830. [45] O. S. Al-Kadi, A multiresolution clinical decision support system based on fractal model design for classi cation of histological brain tumours, Computerized Medical Imaging and Graphics 41 (2015) 67{79. 29 [46] M. E. Mavroforakis, H. V. Georgiou, N. Dimitropoulos, D. Cavouras, S. Theodoridis, Mammographic masses characterization based on local- ized texture and dataset fractal analysis using linear, neural and sup- port vector machine classi ers, Arti cial Intelligence in Medicine 37 (2) (2006) 145{162. [47] K. M. Iftekharuddin, W. Jia, R. Marsh, Fractal analysis of tumor in brain mr images, Machine Vision and Applications 13 (5-6) (2003) 352{ [48] R. Lopes, N. Betrouni, Fractal and multifractal analysis: A review, Medical Image Analysis 13 (4) (2009) 634{649. [49] I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE Transactions on Information Theory 36 (5) (1990) 961{1005. [50] O. S. Al-Kadi, A fractal dimension based optimal wavelet packet analysis technique for classi cation of meningioma brain tumours, in: 2009 16th IEEE International conference on image processing, 2009, pp. 4177{ [51] F. Davnall, C. Yip, G. Ljungqvist, M. Selmi, F. Ng, B. Sanghera, B. Ganeshan, K. Miles, G. Cook, V. Goh, Assessment of tumor het- erogeneity: an emerging imaging tool for clinical practice?, Insights into Imaging 3 (6) (2012) 573{589. [52] D. Mitrea, S. Nedevschi, M. Lupsor, M. Socaciu, R. Badea, Classi cation of the hepatocellular carcinoma in ultrasound images based on the imag- istic textural model of this tumor, in: S. Vlad, R. Ciupa, A. Nicu (Eds.), International Conference on Advancements of Medicine and Health Care through Technology, Vol. 26 of IFMBE Proceedings, Springer Berlin Hei- delberg, 2009, pp. 267{272. [53] H. S. Christopher D. ManningPrabhakar Raghavan, Introduction to In- formation Retrieval, 1st Edition, Cambridge University Press, 2008. [54] P. Domingos, M. Pazzani, On the optimality of the simple bayesian classi er under zero-one loss, Machine Learning 29 (2-3) (1997) 103{ 30 [55] L. Mo, R. Cobbold, A uni ed approach to modeling the backscattered doppler ultrasound from blood, IEEE Transactions on Biomedical En- gineering 39 (5) (1992) 450{461. [56] O. Al-Kadi, Tumour grading and discrimination based on class assign- ment and quantitative texture analysis techniques, Phd thesis, Depart- ment of Informatics (2009). http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Electrical Engineering and Systems Science arXiv (Cornell University)

Heterogeneous tissue characterization using ultrasound: a comparison of fractal analysis backscatter models on liver tumors

Download PDF
31 pages

Loading next page...
 
/lp/arxiv-cornell-university/heterogeneous-tissue-characterization-using-ultrasound-a-comparison-of-7E1lNviNJh

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

ISSN
0301-5629
eISSN
ARCH-3348
DOI
10.1016/j.ultrasmedbio.2016.02.007
Publisher site
See Article on Publisher Site

Abstract

Assessing tumor tissue heterogeneity via ultrasound has recently been sug- gested for predicting early response to treatment. The ultrasound backscat- tering characteristics can assist in better understanding the tumor texture by highlighting local concentration and spatial arrangement of tissue scatter- ers. However, it is challenging to quantify the various tissue heterogeneities ranging from ne-to-coarse of the echo envelope peaks in tumor texture. Lo- cal parametric fractal features extracted via maximum likelihood estimation from ve well-known statistical model families are evaluated for the purpose of ultrasound tissue characterization. The fractal dimension (self-similarity measure) was used to characterize the spatial distribution of scatterers, while the Lacunarity (sparsity measure) was applied to determine scatterer num- ber density. Performance was assessed based on 608 cross-sectional clinical ultrasound radio-frequency images of liver tumors (230 and 378 demonstrat- ing respondent and non-respondent cases, respectively). Cross-validation via leave-one-tumor-out and with di erent k -folds methodologies using a Bayesian classi er were employed for validation. The fractal properties of the backscattered echoes based on the Nakagami model (Nkg) and its extend four-parameter Nakagami-generalized inverse Gaussian (NIG) distribution achieved best results { with nearly similar performance { for characterizing liver tumor tissue. Accuracy, sensitivity and speci city for the Nkg/NIG Corresponding Author: E-mail. o.alkadi@ju.edu.jo. Preprint submitted to Ultrasound in Medicine and Biology December 23, 2019 arXiv:1912.09903v1 [eess.IV] 20 Dec 2019 were: 85.6%/86.3%, 94.0%/96.0%, and 73.0%/71.0%, respectively. Other statistical models, such as the Rician, Rayleigh, and K-distribution were found to not be as e ective in characterizing the subtle changes in tissue texture as an indication of response to treatment. Employing the most rel- evant and practical statistical model could have potential consequences for the design of an early and e ective clinical therapy. Keywords: Ultrasound, tissue characterization, liver tumor, RF envelope, fractal analysis, texture analysis 2 Introduction Ultrasound tissue characterization can provide useful quantitative assess- ments for understanding the state of biological disease [1]. With advancement in medical image analysis, it is becoming a promising non-invasive technique for early detection of tumor response to treatment [2, 3, 4]. It has the ad- vantage of deriving parameters that can represent tissue properties in a fast, non-ionizing, easily operated, and cost-e ective way compared to other con- ventional follow-up imaging techniques. Soft tissue pathologies in the form of lesions tend to have distinct scattering patterns to that of normal tis- sue structure, and the associated acoustic properties could characterize the concentration of scatterers and micro-structures; which is an indication of di erent tissue types. Biological tissue ultrasonic modeling followed by echo signal analysis can facilitate heterogeneity examination of tumor texture. The interaction of an acoustic wave with di erent tissue regions can be modeled by the backscattered radio-frequency (RF) signal. Tissue proper- ties based on the scatterer number density and spatial distribution can be derived subsequently for analysis. There are several approaches for which useful information can be extracted from the RF signal. One approach uses the local power spectral density to estimate the integrated backscatter and attenuation coecient [5, 6, 7, 8], or to measure the mean central frequency and scatterer size [9, 10, 11, 12]. Textural properties of the tissue spatial arrangement can also be estimated from the envelope-detected RF image [13, 14]. As the rst-order statistical properties of the backscattered RF sig- nal rely on the number density and spatial distribution of scatterers [15, 16] { which maybe coherent, random or a mixture of both { it would be dicult to account for all scatterer conditions using the former approaches. Therefore others have investigated the probability density function of the backscattered echoes and proposed to account for the number, size, spacing and regularity of the scatterers in tissue. An overview of the various statistical distributions for modeling the envelope-detected RF signal can be found in [17]. For the latter approach, the main objective is to provide a better characterization of the fundamental elements that form the coarse textural patterns, namely speckles formed from the backscattered echoes. The speckle local arrange- ment represents the various scatterer concentrations and spatial distributions occurring in tissue; ranging from a fully developed, partially developed, to a coherent speckle pattern. In cases when there are many randomly located scatterers per resolution cell (i.e. fully developed speckle), the envelope sig- 3 nal statistics would follow a square root of exponential distribution, known as Rayleigh distribution [15]. The model can be further subdivided into pre-, Rayleigh, and post-Rayleigh for characterizing heterogeneous, homogeneous, and periodic textures, respectively [18, 19, 20]. Non-Rayleigh distributions can be observed when the scatterers become less condensed or having a struc- ture with some regularity. That is, when the number of scatterers in a reso- lution cell is small (i.e. partially developed speckle), the K-distribution was found to be more e ective in modeling the pre-Rayleigh statistics of the RF envelope [21]. The RF envelope statistics in this case would have a shape resembling a square root of the product of a gamma and exponential distribu- tions. Whereas if the scatterers become organized with some periodicity (i.e. coherent speckle), the Rician distribution { which is the generalization of the Rayleigh distribution { is more appropriate for characterizing the underlying regular structures in tissue [22]. Although the prior models account for speci c aspects of scatterer lo- calization, they are not general enough to model the various tissue texture conditions. This is true when a high degree of variability in the scattering cross-section associated with a low number of scatterers exist. In practice it is very common to encounter non-Rayleigh conditions of ultrasound backscat- ter, such as mixtures of di use and coherent or periodically aligned scatterers in tissue micro-structure. Therefore a comprehensive approach is sought as in the generalized K-distribution [23] and homodyne K-distribution [24, 25]. A third parameter, which represents the envelope of the coherent signal, was added to the e ective number of scatterers and energy of the random scatter- ers to account for post-Rayleigh conditions. However a drawback lies in the analytical complexity of these model generalizations, rendering the process of parameter estimation computationally expensive. Other models which also apply a generalization approach to the Rayleigh and Rice distributions to better- t the backscattered echo include: Weibull [26], Rician inverse Gaus- sian [27], and generalized gamma [28]. On the other hand, the Nakagami distribution family can provide a simpler and general model for ultrasonic tissue characterization [29]. The regularity of the scatterer spacing are taken into consideration besides the scatterer density and amplitude, making it pos- sible to account for hypoechoic and hyperechoic structures [30]. The model shape equivalent to a scaled square root of a gamma distribution can better represent the backscattered RF signal envelope, and can be easily ne-tuned via the shape Nakagami parameter to represent low and high number of scat- terer densities with minimal error [31]. The model was further generalized 4 as a Nakagami-generalized inverse Gaussian distribution in [32] by including an additional shape adjustment parameter to account for the tails of the density function. However this generalization was not investigated with real tissue, where scatterers tend to have a high degree of variability in scattering cross-sections. All aforementioned statistical models of the backscattered echo envelope in the literature claim a better characterization of texture anisotropic proper- ties. A large number of articles address the problem of soft tissue characteri- zation and diagnosis from ultrasound images of various internal organs, such as in kidneys [33], liver [34], breast [35], gallbladder [36], pancreas [37], spleen [38] and abdominal aorta [39] all of which demonstrates practical examples, but not limited to, of recent clinical work . However, to the best of our knowledge, the study of the performance of the most well-known statistical models for characterizing ultrasonic liver tumor tissue has not been investi- gated. Local parametric fractal features extracted via maximum likelihood estimation for each statistical model will be used for liver tumor texture clas- si cations. In this paper a fractal approach for evaluating the performance of the di erent statistical models would be sought for an automated detection of liver tumor response to chemotherapy treatment. The approach relates the fractal characteristics of the tissue scatterers to the underlying statis- tical properties derived from the RF envelope-detected signal. The fractal dimension which represents the degree of self-similarity and the derived La- cunarity which indicates the level of spaces within the texture were related to the scatterers spatial distribution and number density, respectively. The fractal features are extracted via a multimodal statistical distribution and results fed to a classi er for automated classi cation. The aim of the paper is to demonstrate the ecacy in modeling the underlying tumor physiologi- cal changes from ultrasound parametric images of scatterer tissue properties. Tumor tissue scatterer characterization is a challenging task due to the dis- ease chaotic behavior. Therefore we will address the issue of best model selection, and whether analytically complex models are really necessary for better characterizing complex tissue, such as liver tumors. The paper is organized as follows: the preference of working with RF envelope-detected signal rather than B-mode images is brie y explained in the ultrasound data representation section, and followed by modeling the backscattered RF signal to cover the di erent scatterer conditions and densi- ties. Then the multiscale feature extraction section discusses characterizing the fractal patterns of the tissue scatterers for automated classi cation. The 5 results and discussion sections list and discuss the major ndings, and the paper is summarized in the conclusion section. Ultrasound data representation RF parametric image acquisition The acquired ultrasound beams run through several signal and image pre-processing steps before they can be presented as a radio-frequency (RF) matrix, and subsequently in a gray scale B-mode image, see Fig.1.The ul- trasound operator may also choose to further optimize the visibility by in- creasing the gain of the re ected signals with increasing time from the trans- mitted pulse (i.e. time-gain compensation). While the B-mode scans could be suitable for qualitative analysis in describing the tissue structure, some aspects could be in fact obscured or modi ed, and would not realistically represent the condition in the raw data. The post-processing functions that have been applied manipulate the image on the screen, without improving the signal quality itself, or the fundamental signal to noise ratio. The RF envelope-detected signal can better provide quantitative analysis and bal- ance the tradeo between preserving most of the ultrasound data unaltered and with minimal ltering (i.e. only removal of high frequency oscillations). Many works in the literature has shown the usefulness of RF data for clinical analysis as mentioned in Section 1, therefore this work utilizes the envelope- detected RF data for ultrasound tissue characterization of liver tumors. Clinical liver tumor cross-sectional images Cross-sectional images of liver tumors undergoing chemotherapy treat- ment obtained as part of an ethically approved prospective study was used to validate the clinical signi cance of the di erent statistical models. A to- tal of 608 cross-sectional images (230 from tumors responded to treatment categorized as respondent acquired from 11 patients, and 378 from tumors progressed categorized as non-respondent acquired from 22 patients) were obtained using a diagnostic ultrasound system (z.one, Zonare Medical Sys- tems, Mountain View, CA, USA) with a 4 MHz curvilinear transducer and 11 MHz sampling. The target tumor was initially imaged using the optimal acoustic window as determined by an experienced operator with 5 years experience of liver ultrasound imaging. The depth of imaging was kept constant for all study acquisitions. Other imaging settings including scanning frequency, gain and 6 Figure 1: Illustration of ultrasound RF to B-mode data conversion process. time gain compensation were also kept constant. The operator then posi- tioned the imaging transducer to ensure the target tumor is optimally placed at the center of the image. While maintaining stable skin contact position, the transducer was angled to image the most cranial aspect of the tumor. The transducer was then swept in a fan motion towards the caudal aspect of the tumor, capturing multiple cross sectional images through the target tumor. The procedure was repeated at all subsequent time points during the moni- toring period. Since the imaging protocol was constant, the tumor should be at identical depth every time it is imaged throughout the monitoring period. Each dataset was acquired prior to commencement of chemotherapy. Re- sponse to treatment was determined based on conventional computed tomog- raphy follow up imaging as part of the patient standard clinical care based on the response evaluation criteria in solid tumors (RECIST) [40]. The base- line cross-sectional imaging was compared against those performed at the end of treatment according to the RECIST criteria to determine response to treatment for each target tumor. A tumor was classi ed as responsive if 7 categorized as partial response and non-responsive if no change or disease demonstrated progression. Cross-validation was performed based on a leave- one-tumor-out approach for a total of 40 volumetric (stacks of segmented 2D slices) tumors, for which 13 were responsive and 27 did not respond to chemotherapy treatment. Imaging was done by two radiologists to provide gold standard and training data. The acquisition of the dataset in this work did not in uence the diagnostic process or the patients' treatment, and in- formed consent was received from each participant in the study. Modeling the backscattered RF signal Ultrasound backscattering in tissue can be considered as a random pro- cess. Many statistical models have been applied for describing the proba- bility density function of the envelope-detected RF signal. Evaluation has been made by ve major statistical distributions that are widely used in the literature to describe the various aspects of tissue characteristics. Rayleigh distribution The Rayleigh distribution ideally suits the condition when the ampli- tudes of the individual backscattered signals are assumed to be randomly distributed, i.e. speckle pattern is \fully developed". The n -dimensional Rayleigh distribution [17] can be seen as a continuous probability distribu- tion for positive-valued random variables de ned n=2 2 1 2 2 2 n1 x =2 P xj = x e ; (1) ray (n=2) 2 where x represents the amplitude of the signal,  > 0 is the scale parameter of the distribution, and denotes the Euler gamma function. The case when n = 2 corresponds to the Rayleigh distribution. Rician distribution The Rician or Rice distribution model combines a coherent component to the backscattered signal that underlies the Rayleigh distribution. It has the capability of modeling unresolved structures exhibiting a periodic pattern in the spatial scatterer distribution [41]. An n -dimensional Rician distribution [22] can be expressed as n=2 x  2 2 2 +x =2 2 ( ) P xj;  =  I x e ; (2) ric n=21 2 2 8 where x represents the amplitude of the signal, the scale parameter  > 0 and non-centrality parameter   0 are real numbers, n is the dimension and I denote the modi ed Bessel function of the rst kind of order p. When n = 2 the distribution would correspond to a Rician distribution.The special case where  ! 0 would yield the Rayleigh distribution. K-distribution The K-distribution can be considered as a generalization of the Rayleigh distribution that can account for non-uniformities in the spatial scatterer distribution. This property allows the modeling of clusters of scatterers or when the e ective number of scatterers per resolution cell is small. The K-distribution [17] is de ned by 1+n=2 4x 2 P xj ; = K x ; (3) kd n=2 ( +n=2)=2 2 (2 ) ( ) (n=2) where > 0,  > 0 are the shape and scale parameters, respectively, and K denotes the modi ed Bessel function of the second kind of order p. The K-distribution includes a Rayleigh distribution as a special case for = 1. Nakagami distribution This is another family of distributions that can model the ultrasonic backscattered envelope for various scattering conditions and scatterer densi- ties. The Nakagami density function [42] is de ned as 2m 2m1 mx = P (xjm; ) = x e ; (4) nkg (m) for x  0, where is the Euler gamma function. The real numbers m > 0 (related to the local backscattered energy) and > 0 (related to the local concentration of scatterers) are called the shape and scaling parameters, respectively. Similarly to the Rayleigh distribution, the envelope of the RF signal x follows a gamma distribution. By ne-tuning the shape of the distribution parameter m, other statistical distributions can be modeled, such as, an approximation of the Rician distribution (i.e. post-Rayleigh) for m > 1, a Rayleigh distribution for the special case when m = 1, and when m < 1 a K-distribution (i.e. post-Rayleigh) for modeling the backscattered echoes with less dense scatterers. Thus the SNR of the Nakagami distribution can take any positive value. 9 Nakagami-generalized inverse Gaussian distribution Based on a four parameter model, the Nakagami-generalized inverse of Gaussian distribution (NIG) distribution [32, 17] is an extended form of the Nakagami distribution to better cover the coherent scattering conditions. It is de ned by 2 (m= 2m1 P (xjm; ; ; ) = x NIG p =2 K  (m) (m)=2 2mx K  + ; (5) 2mx + where m, ,  and  are positive real numbers with the condition  > 0; 0, and K is modi ed Bessel function of the second kind. By having the four parameter values of the NIG distribution model set to speci c conditions, various well-known distributions can be recovered such as, Rayleigh (m = 1;  > 0; ! 1), Nakagami (m > 1;  > 0; ! 1), K-distribution (m = 1;  > 0; ! 0), and Nakagami-gamma (m > 0;  > 0; ! 0). Besides the NIG distribution, other complex models such as the Generalized- K, the Homodyned-K, and the Rician inverse of Gaussian distributions exist in the literature for ultrasound tissue characterization. However they were not used as the simpler and more general models of Nakagami family { in its 2 and 4-parameter extended form { which under di erent limiting condi- tions can approximate the known distributions, and hence can cope with the various scattering conditions related to the coherent-to-di use signal power ratio and scatterer clustering, see illustration in Fig.2. Multiscale feature extraction Parametric image generation Based on a voxel-by-voxel approach, the statistical model parameters are estimated from the RF envelope data. Each estimated parameter in the generated parametric image corresponds to a voxel { short segment { of the RF signal. As we are dealing with a tumor volume, the RF enve- lope signal was presented as a matrix, and several matrices represent the tumor volume. Then the di erent statistical distribution models are t- ted and then their associated model parameters estimated by maximum likelihood estimation. Namely, the maximum likelihood estimate  (v) for 10 Scatterer characterization Figure 2: An illustration representing the coverage of the scatterer distribution vs scat- terer number density by the 5 statistical distributions employed in this work for tissue characterization. a density function f (v ; : : : ; v =) when  is a vector of parameters for 1 m a certain distribution family , estimates the most probable parameters (v) = argmax D (=v ; : : : ; v ), where D (=v) = f (v=) ;  2  is the 1 m score function. Thereby, a number of parametric images l , where r = 1; 2; or 4 for a mono, bi, and quad-parametric model, respectively, are generated for each corresponding RF envelope image of the applied distribution model. Herein, l refers to the statistical model parameters estimated locally from the neighborhood of the RF envelope data by a square{sliding window with its size optimized as in [43] and centered on each point in the RF data matrix. Also the same RF envelope data was used to create the di erent parametric images for each statistical model. That is, creating for each model maps of the distribution parameters throughout the entire scanned region (i.e. each parameter map being an image). An interesting point to consider is the di erence between the spatial dis- tribution of tissue scatterers, and the structure of the texture of the para- metric ultrasound images. The former results from the interaction of the 11 sound energy with the liver tumor tissue leading to scatter in many direc- tions. Complex structure of tissue scatterers and their small size as com- pared to the ultrasound wavelength, gives the speckle patterns a Rayleigh scattering behavior. The spatial distribution of these tissue scatterers is not related to any particular ultrasound imaging system, but is a property of the tissue structure and the ultrasound frequency used which dictates the scat- tering length scale. On the other hand, texture of the parametric ultrasound images, which shows di erent spatial variations of speckle intensity, is depen- dent on the used ultrasound imaging system. Usually a set of proprietary lters { speci c to the ultrasound device manufacturer { are used for post- processing operations, which may result in loss of information. Therefore the RF envelope-detected data in this work did not undergo any post-processing ltering or time-gain compensation operations. Fractal features Various statistical signatures can be derived from the acoustic properties of the backscattered RF and envelope-detected signals. An e ective way to investigate the local arrangement of the scatterer concentrations and spatial distributions occurring in tissue is using a multifractal analysis approach. The fractal dimension (FD) can indicate the degree of \self-similarity" in the spatial distribution of tissue scatterers, and has shown many promising results in tissue characterization from di erent imaging modalities [44, 45, 43, 46, 47]. Also the Lacunarity, which is a measure derived from the texture of the FD, re ects the gaps within the tissue texture, and hence measures the number density of tissue scatterers. Both signatures, the FD de ned in (6) { estimated via the fractal Brownian motion which is a non-stationary model known for its capability for describing random phenomena [48] { and its counterpart Lacunarity which is the standard deviation divided by the mean of a fractal image F using a sliding box-counting algorithm as shown in (7), are considered useful for revealing the mixtures of spatial homogeneity and heterogeneity properties of tissue texture in the generated parametric images. log (N ) FD = ; (6) log (1=s) P P M1 N1 2 1=MN F (m; n) m=0 n=0 L = 1 (7) P P M1 N1 1=MN F (k; l) k=0 l=0 12 where N is the number of self-similar shapes and s is the corresponding scaling factor; M and N in (7) are the size of the fractal parametric image F . In this work, the FD and derived Lacunarity signatures are estimated in a multiscale fashion after applying a Daubechies wavelet packet analysis de- composition technique, where the texture variations of the subbands at each level of decomposition serve as the extracted fractal feature pattern. The Daubechies wavelet can account for self-similarity and signal discontinuities, making it useful for characterizing signals exhibiting fractal patterns [49]. An orthogonal 8-tap Daubechies lter was used to obtain the wavelet packets by expanding the basis having the most signi cant fractal signature rather than energy [50]. The fractal features at the di erent resolution scales give more information on the local concentration and spatial arrangement of the tissue scatterers { as higher FD signature values indicate a more heterogeneous texture (i.e. distribution of scatterers are random), and vice versa [51, 52]; whereas Lacunarity signature values operate oppositely, giving a low value when the scatterers are large, viz. less gaps between scatterers per resolution cell, see Fig.3. An illustrative example of liver tumor tissue characterizing by the di erent statistical models is shown in Fig.4. The rst column shows the distribution and number density of the tissue scatterers modeled by the di er- ent statistical distributions, while the second column represents the textural patterns of the modeled tissue scatterers in a multifractal analysis approach. l l At the end of the feature extraction stage, a feature vector = = (w ; w ; 1;1 1;2 : : : ; w ) would consist of all selected fractal signatures w for a speci c model r;q parameter r and wavelet subband q from each RF envelope image l. A data pipeline analysis of the fractal feature extraction process is illustrated in Fig.5. Pattern classi cation In order to assess the quality of the extracted features from the di erent statistical distribution models, a na ve Bayesian classi er (NBC) was used for classi cation. The NBC is a robust learning method which can work well even when the data is non-normally distributed [53], and the assumption of attribute independence allows the NBC classi er to estimate the classi cation accuracy using less training data as compared to other classi ers. This fast and simple classi er was reported to perform well in practice even with the presence of strong attribute dependence [54]. According to Bayesian theory, 13 Fractal characterization Figure 3: Tissue characterization using the fractal properties of the scatterers distribution and clustering: The fractal dimension measures the level of self-similarity (i.e. spatial distribution), while the Lacunarity assesses the degree of heterogeneity (i.e. denseness) in tissue texture. each fractal feature = ; f = 1; : : : ; n, would be assigned to class i, if P (C j=) f i is maximized over all classes P (C j=) > P (C j=)8j 6= i. i j To nd P (C j=), NBC can be represented as: P (C ) P (= jC ) i f i f =1 P (C j=) = (8) P Q k n P (C ) P (= jC ) i f i i=1 f =1 where P (C j=) is the a posteriori probability of assigning class i given feature th vector =, P (= jC ) is the probability density function of = within the i f i class C for a total number of k classes, P (C ) and P (=) are the a priori i i probability of class C and feature vector =; respectively. Results The extracted fractal features from the texture of the various parametric images { which where derived from the RF envelope-detected signal of the dif- ferent distribution models {are compared in Tables 1, 2, and 3, where values 14 Figure 4: Characterizing a segment of a liver tumor tissue: (from left to right) model para- metric image and corresponding fractal image for (a) Nakagami, (b) Nakagami-generalized inverse Gaussian, (c) Rayleigh, (d) Rician, and (e) K-distribution statistical models, re- spectively. 15 Figure 5: Data pipeline analysis explaining the fractal feature extraction process. 16 in bold indicate the best achieved performance. A detailed classi cation per- formance of the 3D clinical ultrasound liver tumor test set via a leave-one-out cross-validation approach is shown in Table 1. The tissue characterization based on the Nakagami (Nkg) distribution and its 4-parameter extended ver- sion (NIG distribution) outperformed the other statistical models in terms of classi cation performance. Both Nkg and NIG had nearly equal classi ca- tion accuracy and the area under the receiver operating characteristics (ROC) curve with 85.6% & 86.3%, and 83.2% & 83.5%, respectively. Although NIG had the highest sensitivity (true-positive rate), the Nkg achieved relatively better in the statistical measures of speci city (true-negative rate), precision, and false-positive (FP) rate. The classi cation performance of the Rayleigh distribution was the third among all ve distributions; nevertheless achiev- ing equal statistical measures regarding speci city and FP rate with the NIG model. The Dice similarity coecient arms the accuracy results and shows how similar the predicted and observed responses are in regard to chemother- apy treatment. Although the K-distribution was the least sensitive in terms of classi cation performance { next to Rician distribution, it recorded the best speci city score as compared to the other distributions. In order to avoid over tting and to give some insights on how would the di erent distribution models based on the scatterers fractal characteristics generalize, the extracted features were further validated. Table 2 and 3 show the detailed classi er performance on the test set for 10- and 20-fold cross- validation (results are the mean  standard deviation of the performance over 60 runs), respectively. Table 1: Detailed Classi cation Performance (leave-one-out cross-validation) for di erent statistical models of RF ultrasound for liver tissue characterization Classi cation RF-based stochastic distribution Performance Nkg Ric Ray NIG Kd FP rate 0.271 0.327 0.289 0.289 0.336 Sensitivity 0.940 0.890 0.920 0.960 0.880 Speci city 0.730 0.670 0.710 0.710 0.770 Accuracy 0.856 0.810 0.838 0.863 0.798 Precision 0.850 0.810 0.830 0.840 0.795 Dice SC 0.922 0.892 0.912 0.926 0.888 ROC Area 0.832 0.781 0.814 0.835 0.773 17 Table 2: Detailed Classi cation Performance (10-fold cross-validation) for di erent statistical models of RF ultrasound for liver tissue characterization Classi cation RF-based stochastic distribution Performance Nkg Ric Ray NIG Kd FP rate 0.280  0.076 0.345  0.113 0.285  0.089 0.332  0.045 0.380  0.128 Sensitivity 0.924  0.720 0.887  0.655 0.911  0.715 0.955  0.668 0.870  0.619 Speci city 0.720  0.924 0.655  0.887 0.715  0.911 0.668  0.955 0.620  0.872 Accuracy 0.845  0.007 0.797  0.009 0.836  0.007 0.845  0.009 0.780  0.011 Precision 0.840  0.856 0.803  0.785 0.836  0.836 0.821  0.905 0.790  0.754 Dice SC 0.916  0.004 0.887  0.005 0.910  0.004 0.916  0.006 0.870  0.007 ROC Area 0.822  0.009 0.771  0.010 0.813  0.008 0.812  0.011 0.75  0.012 Table 3: Detailed Classi cation Performance (20-fold cross-validation) for di erent statistical models of RF ultrasound for liver tissue characterization Classi cation RF-based stochastic distribution Performance Nkg Ric Ray NIG Kd FP rate 0.274  0.073 0.332  0.112 0.283  0.084 0.308  0.041 0.361  0.124 Sensitivity 0.927  0.726 0.888  0.668 0.916  0.717 0.959  0.692 0.876  0.639 Speci city 0.726  0.927 0.668  0.888 0.717  0.916 0.692  0.959 0.639  0.876 Accuracy 0.849  0.007 0.803  0.005 0.839  0.007 0.856  0.008 0.785  0.008 Precision 0.843  0.862 0.809  0.789 0.837  0.843 0.832  0.915 0.794  0.765 Dice SC 0.918  0.004 0.891  0.003 0.912  0.004 0.922  0.004 0.879  0.005 ROC Area 0.826  0.007 0.778  0.006 0.816  0.008 0.826  0.009 0.758  0.010 Since we are concerned more with the ability of an early chemotherapy treatment to identify initial tissue response to treatment, namely the sensi- tivity of the treatment procedure, Fig.6 shows the di erence in the sensitivity measure between the raw RF parameters and the fractal-based measures ex- tracted from the parametric images of the RF envelope detected signal. The Wilcoxon signed-rank test on paired overall sensitivity for the fractal and parametric-based of the statistical models shows there is signi cant di er- ences between the two approaches (p < 0:05). For completeness, the models accuracy are shown in Fig.7 as well. 18 Sensitivity Comparison Figure 6: Sensitivity comparison of the di erent statistical models for tissue character- ization based on the using the estimated model parameters (parametric-based) and by the fractal properties of the tissue scatterers(fractal-based). The models sensitivity was cross-validated via (a) leave-one-out, (b) 10-fold, and (c) 20-fold. Discussion Heterogeneity in the tumor tissue scatterers could span di erent scatter- ing conditions. Regions within the tumor tissue which respond to treatment 19 Accuracy Comparison Figure 7: Cross-validation accuracy comparison of the di erent statistical models for tissue characterization based on the using the estimated model parameters (parametric-based) and by the fractal properties of the tissue scatterers(fractal-based). The models accuracy was cross-validated via (a) leave-one-out, (b) 10-fold, and (c) 20-fold. might exhibit di erent statistical properties to that of the non-respondent counterpart. Thus it is essential to evaluate the discriminative abilities of 20 the statistical models while simultaneously taking into account all possible scatterer conditions and densities. The sum of the individual backscattered signals from the various scatter- ing regions within the biological tissue can be used for RF signal modeling. In the case of a fully developed speckle pattern, the size of the scatterers tend to be small in comparison with the wavelength and also have a disor- ganized structure, i.e. the spatial distribution of scatterers is random [55]. This can be considered as a typical case for tumor tissue, where the tumor angiogenesis process causes for the newly induced blood vessels to grow in a chaotic manner, and therefore tissue scatterers to have a random nature. The relation between angiogenesis and randomness of tissue scatterers can be attributed to the chaotic way that tumors build their network of new blood vessels, and thus giving tumorous tissue a \rougher" appearance than non-tumorous. These angiogenesis networks tend to be leaky and disorga- nized, unlike blood vessels in normal tissue. The chaotic behavior introduces a degree of randomness in appearance or \roughness" and gives a higher fractal dimension value compared to necrotic tissue. However spatial scat- terer distribution in soft tissue is not strictly random and may not originate from identical scatterers, also it is common to encounter di erent degree of variability in tumor tissue as scatterers vary in shape and size, with a mix- ture of partially developed and coherent speckle patterns. This compound situation becomes more distinct once certain regions within the tumor start to respond to treatment, which re ects in the statistics of the RF backscat- tered signal. Moreover, after generating the parametric images from the RF envelope echo, the fractal characteristics of tissue scatterers can provide addi- tional information on the local spatial distribution and number density. Fig.6 and Fig.7 illustrate an improved performance when a fractal characterization is applied to the parametric images. The fractal dimension is analogous to the coherent-di use ratio in Fig.2, where it tends to improve distribution separability from a texture perspective by assessing the scatterers random- periodic properties. Also the Lacunarity of the textures in the parametric images further re nes the scatterer densities and measures the degree of het- erogeneity in tumor tissue. The fractal approach shows a better sensitivity for all distribution models in Fig.6 with the best improvement achieved by the Nakagami model, while the NIG model recorded the highest sensitiv- ity. Similarly in Fig.7, where the accuracy is remarkably improved for the fractal-based images of the Nakagami model, and achieving a nearly compa- rable performance to that of the complex NIG model. A point to note is the 21 decrease in the K-distribution accuracy for the cross-validations in Fig.7(b) and (c). This indicates that the fractal approach is incapable of extracting useful information from the tissue scatterers characterized by the model, and hence giving inaccurate characterization of tumor heterogeneity. In some typical conditions of ultrasound tissue characterization, the backscat- tered signal cannot be assumed to be Rayleigh distributed, such as when the number of scatterers per resolution cell are not large enough or when there is a lack of randomness in the scatterers localization (e.g. due to period- icity, structure, or clustering) [41]. The former conditions suggest that the Rayleigh distribution may not approximate e ectively the statistical charac- teristics of the echo envelope when tumor tissue regions becomes less dense and having a more consistent structure. Although tumor tissue is known to have a chaotic nature [56], the change in the tissue scatterer textural proper- ties in a sense of becoming less random (e.g. inducing much less angiogenesis), or even the possible reduction in the number of e ective scatterers in some parts of the tumor tissue (e.g. more necrotic regions) could be attributed to initial response to treatment and tumor becoming less aggressive. This explains why the Rayleigh distribution in Table 1 did not perform as well as the Nakagami distribution family which can take into account such variabil- ity in both tissue scatterer distribution and number density. On the other hand, it is not expected to encounter periodic patterns within the tumor tissue due to its chaotic nature, therefore dominance of periodic or coher- ent patterns in the modeled tissue texture is very rare, which explains the reduced classi cation performance based on the Rician distribution. As a result, the spatial scatterer distribution non-uniformities, such as clusterings in scatterers or regions with smaller e ective numbers of scatterers cannot be e ectively modeled. Also the K-distribution can be used to model the spacing or regularity in the spatial scatterer distribution, such as in partially developed speckle based on a low-average number of e ective scatterers per resolution cell; however, it had the least classi cation performance in terms of predicting liver tumor response to chemotherapy treatment. This can be seen as tissue scatterers in liver tumors tend to have a high degree of num- ber density which the model does not cover, and relying on characterizing the tissue randomness only might not suce. Also the results in Table 1, 2, and 3 show the NIG recorded nearly similar performance in terms of clas- si cation accuracy to that of the Nakagami model. The main advantage of the extended NIG model over the bi-parameter Nakagami model is the bet- ter coverage of the coherent-di use region for low scatterer number density. 22 Since there is not much obvious structural or periodic components in the tissue echo signal of the liver tumor, both models exhibited similar behavior. Moreover, the higher sensitivity values for the NIG model suggest that it is better in predicting response to chemotherapy treatment (i.e. aggressive tu- mors that are responding to treatment), while the Nakagami model is more precise and has the least overall FP rate. Finally, the relatively low digital sampling rate could be a potential lim- itation. The e ect of sparse sampling of the analog signal on the fractal analysis of the RF signal needs to be further investigated. Also as tumors may be localized at varying depths, this may decrease the amplitude of the RF data. Therefore the performance of the derived model parameters based on the fractal analysis technique as a function of depth could serve as future work. Furthermore, the study of the link between the FD derived from para- metric images of weak scattering sources to the spatial distribution of tissue scatterers would assist in better understanding the robustness of the fractal analysis approach. Conclusions This paper presents an evaluation of di erent statistical models that cover di erent aspects of the tissue scatterers distributions and densities. The aim is to understand which model can give a better- t of the complex nature of liver tumors by providing discriminative features for supporting clinical diagnosis as part of healthcare delivery. The tumor heterogeneity was as- sessed via its tissue fractal properties derived from the RF envelope-detected signal and used as features for assessing classi cation performance. The Nak- agami distribution family was the most e ective in characterizing the liver tumor tissue, with the simple bi-parametric model achieving equal accuracy and better precision as compared with the complex 4-parameter extended version. Moreover, fractal analysis of the echo signal was found to give ad- ditional information about the underlying tissue structure, and hence better indicates response to chemotherapy treatment. These ndings suggest that the simple and general models of the Nakagami distribution family along with the fractal analysis of the RF parametric images have promising potential in providing improved ultrasound tissue characterization of liver tumors. 23 Acknowledgements This work was support by the Engineering and Physical Sciences Research Council and Wellcome Trust Grant WT 088877/z/09/z. The authors would like to thank the anonymous reviewers for their constructive comments and suggestions to improve the quality of the paper. 24 References [1] J. Mamou, M. L. Oelze, Quantitative Ultrasound in Soft Tissues, Springer, 2013. [2] G. Czarnota, A. Sadeghi-Naini, N. Papanicolau, O. Falou, R. Dent, S. Verma, M. Trudeau, J.-F. Boileau, J. Spayne, S. Iradji, E. Sofroni, J. Lee, S. Lemon-Wong, M. Ya e, M. Kolios, Quantitative ultrasound evaluation of tumor cell death response in locally advanced breast cancer patients to chemotherapy treatment administration, The Journal of the Acoustical Society of America 133 (5) (2013) 3539. [3] A. Sadeghi-Naini, O. Falou, G. J. Czarnota, Characterizing tumor het- erogeneous response to chemotherapy using low-frequency ultrasonic spectroscopy, The Journal of the Acoustical Society of America 133 (5) (2013) 3584. [4] A. Sadeghi-Naini, N. Papanicolau, O. Falou, J. Zubovits, R. Dent, S. Verma, M. Trudeau, J. F. Boileau, J. Spayne, S. Iradji, E. Sofroni, J. Lee, S. Lemon-Wong, M. Ya e, M. C. Kolios, G. J. Czarnota, Quan- titative ultrasound evaluation of tumor cell death response in locally advanced breast cancer patients receiving chemotherapy, Clinical Can- cer Research 19 (8) (2013) 2163{2174. [5] L. Sannachi, H. Tadayyon, A. Sadeghi-Naini, W. Tran, S. Gandhi, F. Wright, M. Oelze, G. Czarnota, Non-invasive evaluation of breast can- cer response to chemotherapy using quantitative ultrasonic backscatter parameters, Medical Image Analysis 20 (1) (2015) 224{236. [6] N. Rubert, T. Varghese, Scatterer number density considerations in ref- erence phantom-based attenuation estimation, Ultrasound in Medicine and Biology 40 (7) (2014) 1680{1696. [7] K. Nam, J. A. Zagzebski, T. J. Hall, Simultaneous backscatter and attenuation estimation using a least squares method with constraints, Ultrasound in Medicine and Biology 37 (12) (2011) 2096{2104. [8] B. Banihashemi, R. Vlad, B. Debejevic, A. Giles, M. C. Kolios, G. J. Czarnota, Ultrasound imaging of apoptosis in tumor response: Novel preclinical monitoring of photodynamic therapy e ects, Cancer Research 68 (20) (2008) 8590{8596. 25 [9] S. L. Bridal, P. Fornes, P. Bruneval, G. Berger, Parametric (integrated backscatter and attenuation) images constructed using backscattered radio frequency signals (25-56 MHz) from human aortae in vitro, Ultra- sound in Medicine and Biology 23 (2) (1997) 215{229. [10] E. P. Nordberg, T. J. Hall, E ective scatterer diameter estimates for broad scatterer size distributions, Ultrasonic imaging 37 (1) (2015) 3{ [11] R. Lavarello, M. Oelze, Quantitative ultrasound estimates from popula- tions of scatterers with continuous size distributions: E ects of the size estimator algorithm, IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control 59 (9) (2012) 2066{2076. [12] R. K. Saha, M. C. Kolios, E ects of cell spatial organization and size distribution on ultrasound backscattering, IEEE Transactions on Ultra- sonics Ferroelectrics and Frequency Control 58 (10) (2011) 2118{2131. [13] N. Bouhlel, S. Sevestre-Ghalila, Nakagami markov random eld as tex- ture model for ultrasound rf envelope image, Computers in Biology and Medicine 39 (6) (2009) 535{544. [14] T. Klein, M. Hansson, N. Navab, Spatial statistics based feature descrip- tor for rf ultrasound data, 2011 8th IEEE International Symposium on Biomedical Imaging: From Nano to Macro (2011) 33{36. [15] R. F. Wagner, S. W. Smith, J. M. Sandrik, H. Lopez, Statistics of speckle in ultrasound b-scans, IEEE Transactions on Sonics and Ultrasonics 30 (3) (1983) 156{163. [16] M. Pereyra, H. Batatia, Modeling ultrasound echoes in skin tissues us- ing symmetric alpha-stable processes, IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control 59 (1) (2012) 60{72. [17] F. Destrempes, G. Cloutier, A critical review and uniformized represen- tation of statistical distributions modeling the ultrasound echo envelope, Ultrasound in Medicine and Biology 36 (7) (2010) 1037{1051. [18] R. M. Cramblitt, K. J. Parker, Generation of non-rayleigh speckle dis- tributions using marked regularity models, IEEE Transactions on Ultra- sonics Ferroelectrics and Frequency Control 46 (4) (1999) 867{874. 26 [19] P. M. Shankar, R. Molthen, V. M. Narayanan, J. M. Reid, V. Genis, F. Forsberg, C. W. Piccoli, A. E. Lindenmayer, B. B. Goldberg, Studies on the use of non-rayleigh statistics for ultrasonic tissue characterization, Ultrasound in Medicine and Biology 22 (7) (1996) 873{882. [20] R. C. Molthen, P. M. Shankar, J. M. Reid, Characterization of ultra- sonic b-scans using non-rayleigh statistics, Ultrasound in Medicine and Biology 21 (2) (1995) 161{170. [21] P. M. Shankar, V. A. Dumane, J. M. Reid, V. Genis, F. Forsberg, C. W. Piccoli, B. B. Goldberg, Use of the k-distribution for classi cation of breast masses, Ultrasound in Medicine and Biology 26 (9) (2000) 1503{ [22] J. Sijbers, A. J. den Dekker, P. Scheunders, D. Van Dyck, Maximum- likelihood estimation of rician distribution parameters, IEEE Transac- tions on Medical Imaging 17 (3) (1998) 357{361. [23] E. Jakeman, R. J. A. Tough, Generalized k distribution: a statistical model for weak scattering, J. Opt. Soc. Am. A 4 (9) (1987) 1764{1772. [24] V. Dutt, J. F. Greenleaf, Ultrasound echo envelope analysis using a ho- modyned k-distribution signal model, Ultrasonic Imaging 16 (4) (1994) 265{287. [25] J. Mamou, A. Coron, M. L. Oelze, E. Saegusa-Beecroft, M. Hata, P. Lee, J. Machi, E. Yanagihara, P. Laugier, E. J. Feleppa, Three-dimensional high-frequency backscatter and envelope quanti cation of cancerous hu- man lymph nodes, Ultrasound in Medicine and Biology 37 (3) (2011) 345{357. [26] B. I. Raju, M. A. Srinivasan, Statistics of envelope of high-frequency ultrasonic backscatter from human skin in vivo, IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control 49 (7) (2002) 871{882. [27] T. Eltoft, The rician inverse gaussian distribution: A new model for non-rayleigh signal amplitude statistics, IEEE Transactions on Image Processing 14 (11) (2005) 1722{1735. 27 [28] A. S. Tunis, G. J. Czarnota, A. Giles, M. D. Sherar, J. W. Hunt, M. C. Kolios, Monitoring structural changes in cells with high-frequency ul- trasound signal statistics, Ultrasound in Medicine and Biology 31 (8) (2005) 1041{1049. [29] P. M. Shankar, A general statistical model for ultrasonic backscatter- ing from tissues, IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control 47 (3) (2000) 727{736. [30] P. H. Tsui, C. K. Yeh, Y. Y. Liao, C. C. Chang, W. H. Kuo, K. J. Chang, C. N. Chen, Ultrasonic nakagami imaging: a strategy to visualize the scatterer properties of benign and malignant breast tumors, Ultrasound in Medicine and Biology 36 (2) (2010) 209{217. [31] P. H. Tsui, H. Y. Ma, Z. H. Zhou, M. C. Ho, Y. H. Lee, Window- modulated compounding nakagami imaging for ultrasound tissue char- acterization, Ultrasonics 54 (6) (2014) 1448{1459. [32] Karmeshu, R. Agrawal, Study of ultrasonic echo envelope based on nakagami-inverse gaussian distribution, Ultrasound in Medicine and Bi- ology 32 (3) (2006) 371{376. [33] H. P. Wu, N. G. Rognin, T. M. Krupka, L. Solorio, H. Yoshiara, G. Guenette, C. Sanders, N. Kamiyama, A. A. Exner, Acoustic charac- terization and pharmacokinetic analyses of new nanobubble ultrasound contrast agents, Ultrasound in Medicine and Biology 39 (11) (2013) 2137{2146. [34] G. Ghoshal, R. J. Lavarello, J. P. Kemmerer, R. J. Miller, M. L. Oelze, Ex vivo study of quantitative ultrasound parameters in fatty rabbit liv- ers, Ultrasound in Medicine and Biology 38 (12) (2012) 2238{2248. [35] H. Tadayyon, A. Sadeghi-Naini, L. Wirtzfeld, F. C. Wright, G. Czarnota, Quantitative ultrasound characterization of locally advanced breast can- cer by estimation of its scatterer properties, Medical Physics 41 (1). [36] R. E. Kumon, M. J. Pollack, A. L. Faulx, K. Olowe, F. T. Farooq, V. K. Chen, Y. Zhou, R. C. K. Wong, G. A. Isenberg, M. V. Sivak, A. Chak, C. X. Deng, In vivo characterization of pancreatic and lymph node tissue by using eus spectrum analysis: a validation study, Gastrointestinal Endoscopy 71 (1) (2010) 53{63. 28 [37] G. Atiee, H. Kvitko-White, K. Spaulding, M. Johnson, Ultrasonographic appearance of histoplasmosis identi ed in the spleen in 15 cats, Veteri- nary Radiology and Ultrasound 55 (3) (2014) 310{314. [38] B. Roosens, G. Bala, S. Droogmans, J. Hostens, J. Somja, E. Delvenne, J. Schiettecatte, P. Delvenne, V. Caveliers, T. Lahoutte, G. Van Camp, B. Cosyns, Echocardiographic integrated backscatter for assessing re- duction of aortic valve calci cations by r-568 in a rat model of chronic kidney disease, Ultrasound in Medicine and Biology 39 (11) (2013) 2075{ [39] P.-H. Tsui, C.-K. Yeh, C.-C. Chang, Noise e ect on the performance of nakagami image in ultrasound tissue characterization, Journal of Medi- cal and Biological Engineering 28 (4) (2008) 197{202. [40] E. A. Eisenhauer, P. Therasse, J. Bogaerts, L. H. Schwartz, D. Sargent, R. Ford, J. Dancey, S. Arbuck, S. Gwyther, M. Mooney, L. Rubinstein, L. Shankar, L. Dodd, R. Kaplan, D. Lacombe, J. Verweij, New response evaluation criteria in solid tumours: Revised recist guideline (version 1.1), European Journal of Cancer 45 (2) (2009) 228{247. [41] U. Scheipers, Ultrasonic tissue characterization, in: O. Ukimura, I. Gill (Eds.), Contemporary Interventional Ultrasonography in Urol- ogy, Springer London, 2009, pp. 143{177. [42] M. Nakagami, The m-distribution { A general formula of intensity dis- tribution of rapid fading, in: W. C. Ho mann (Ed.), Statistical Methods in Radio Wave Propagation, Pergamon, Elmsford, NY, 1960, pp. 3{36. [43] O. S. Al-Kadi, D. Y. Chung, R. C. Carlisle, C. C. Coussios, J. A. Noble, Quanti cation of ultrasonic texture intra-heterogeneity via volumetric stochastic modeling for tissue characterization, Medical Image Analysis 21 (1) (2015) 59{71. [44] O. S. Al-Kadi, D. Watson, Texture analysis of aggressive and nonag- gressive lung tumor CE CT images, IEEE Transactions on Biomedical Engineering 55 (7) (2008) 1822{1830. [45] O. S. Al-Kadi, A multiresolution clinical decision support system based on fractal model design for classi cation of histological brain tumours, Computerized Medical Imaging and Graphics 41 (2015) 67{79. 29 [46] M. E. Mavroforakis, H. V. Georgiou, N. Dimitropoulos, D. Cavouras, S. Theodoridis, Mammographic masses characterization based on local- ized texture and dataset fractal analysis using linear, neural and sup- port vector machine classi ers, Arti cial Intelligence in Medicine 37 (2) (2006) 145{162. [47] K. M. Iftekharuddin, W. Jia, R. Marsh, Fractal analysis of tumor in brain mr images, Machine Vision and Applications 13 (5-6) (2003) 352{ [48] R. Lopes, N. Betrouni, Fractal and multifractal analysis: A review, Medical Image Analysis 13 (4) (2009) 634{649. [49] I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE Transactions on Information Theory 36 (5) (1990) 961{1005. [50] O. S. Al-Kadi, A fractal dimension based optimal wavelet packet analysis technique for classi cation of meningioma brain tumours, in: 2009 16th IEEE International conference on image processing, 2009, pp. 4177{ [51] F. Davnall, C. Yip, G. Ljungqvist, M. Selmi, F. Ng, B. Sanghera, B. Ganeshan, K. Miles, G. Cook, V. Goh, Assessment of tumor het- erogeneity: an emerging imaging tool for clinical practice?, Insights into Imaging 3 (6) (2012) 573{589. [52] D. Mitrea, S. Nedevschi, M. Lupsor, M. Socaciu, R. Badea, Classi cation of the hepatocellular carcinoma in ultrasound images based on the imag- istic textural model of this tumor, in: S. Vlad, R. Ciupa, A. Nicu (Eds.), International Conference on Advancements of Medicine and Health Care through Technology, Vol. 26 of IFMBE Proceedings, Springer Berlin Hei- delberg, 2009, pp. 267{272. [53] H. S. Christopher D. ManningPrabhakar Raghavan, Introduction to In- formation Retrieval, 1st Edition, Cambridge University Press, 2008. [54] P. Domingos, M. Pazzani, On the optimality of the simple bayesian classi er under zero-one loss, Machine Learning 29 (2-3) (1997) 103{ 30 [55] L. Mo, R. Cobbold, A uni ed approach to modeling the backscattered doppler ultrasound from blood, IEEE Transactions on Biomedical En- gineering 39 (5) (1992) 450{461. [56] O. Al-Kadi, Tumour grading and discrimination based on class assign- ment and quantitative texture analysis techniques, Phd thesis, Depart- ment of Informatics (2009).

Journal

Electrical Engineering and Systems Science – arXiv (Cornell University)

Published: Dec 20, 2019

There are no references for this article.