- Email: [email protected]
Histology-Based Simulations of Ultrasound Imaging: Methodology
Ultrasound in Med. & Biol., Vol. 39, No. 10, pp. 1925–1929, 2013 Copyright Ó 2013 World Federation for Ultrasound in Medicine & Biology Printed in the USA. All rights reserved 0301-5629/$ - see front matter
http://dx.doi.org/10.1016/j.ultrasmedbio.2013.05.005
d
Technical Note HISTOLOGY-BASED SIMULATIONS OF ULTRASOUND IMAGING: METHODOLOGY y x RA SZALAI,z and IMRE KALL ONGY,* LAJOS BALOGH, KLA O* MIKLOS GY€
* Faculty of Information Technology, Pazmany Peter Catholic University, Budapest, Hungary; y National Research Institute for Radiobiology and Radiohygiene, Budapest, Hungary; z Department of Dermatology, Dermatooncology and Venerology, Semmelweis University, Budapest, Hungary; and x Institute of Experimental Medicine, Hungarian Academy of Sciences, Budapest, Hungary (Received 20 November 2012; revised 2 May 2013; in final form 11 May 2013)
Abstract—Simulations of ultrasound (US) images based on histology may shed light on the process by which microscopic tissue features translate to a US image and may enable predictions of feature detectability as a function of US system parameters. This technical note describes how whole-slide hematoxylin and eosin-stained histology images can be used to generate maps of fractional change in bulk modulus, whose convolution with the impulse response of the US system yields simulated US images. The method is illustrated by two canine mastocytoma histology images, one with and the other without signs of intra-operative hemorrhaging. Quantitative comparisons of the envelope statistics with corresponding clinical US images provide preliminary validation of the method. (E-mail: gyongy. [email protected]) Ó 2013 World Federation for Ultrasound in Medicine & Biology. Key Words: Simulation, Modeling, Histology, Envelope statistics, Tissue microstructure, Ultrasonic tissue characterization.
processes such as apoptosis (Vlad et al. 2010). By locating the nuclei in histology slides and calculating certain microstructural statistics from these, a stochastic model has been used to simulate B-mode images whose first-order statistics match those of the corresponding, experimentally obtained images (Daoud and Lacefield 2009). In an alternative, continuum-based approach, 3-D acoustic maps were generated from histology slides and were used to estimate the effective acoustic scatterer diameter of various tissue types (Dapore et al. 2011). Neither model is naturally extensible to modeling higher-order features such as interfaces between tissue types, because only aggregate descriptors are obtained from limited-view histology slices. The current technical note expands on the above work by directly simulating US images from whole-slide histology and comparing their envelope statistics with those of clinical US images. It is hoped that the method described could form the basis of an experimentally validated histology-based model of US image formation.
INTRODUCTION A sufficiently accurate model of ultrasound (US) image formation from microscopic images of tissue could help explain how millimeter-scale features of clinical US images arise from micrometer cellular structures. The model could also be used to predict the viability of an US diagnostic method on a given transducer before clinical testing. Examples of applications are B-mode images of malignant tumors that usually (Catalano et al. 2009), but not always (Hung et al. 2001), appear hypoechoic, and quantitative US, where parameters derived from backscattered US (e.g., envelope statistics) promise better differential diagnosis (Oelze et al. 2007; Tsui et al. 2010). Theoretical models of US image formation abound in the literature, either assuming discrete scatterers or continuously varying acoustic maps (Mari et al. 2009). The discrete scatterer approach has been used to model tissue scattering as arising from clusters of cells (Doyle et al. 2009; Saha and Kolios 2011), helping to explain observed changes in US backscatter resulting from
THEORY OF IMAGE FORMATION Address correspondence to: Miklos Gy€ongy, Faculty of Information Technology, Pazmany Peter Catholic University, Prater utca 50/a, Budapest H-1083, Hungary. E-mail: [email protected]
Using common assumptions (small fluctuations from the mean speed of sound c0 and mean density r0, 1925
1926
Ultrasound in Medicine and Biology
single scattering) found in other continuum scattering models (Cobbold 2006; Jensen 1991), Mari et al. (2009) reported that scattering can be expressed as a function of the longitudinal bulk modulus K only. With suitable rearrangement of terms and a transformation between time t and axial distance z 5 c0t/2, the image u(r) as a function of mapped location r can be expressed as the convolution of an imager response I(r) with a tissue response R(r): uðrÞ 5 IðrÞ RðrÞ
(1)
IðrÞ 5 vpe ðrÞ hpe ðrÞ
(2)
where vpe(r) is the spatial-domain pulse-echo wavelet that includes the electrical excitation, as well as transmit and receive impulse responses, of the transducer (Stepanishen 1981); hpe(r) is the pulse-echo spatial impulse response of the transducer that incorporates transducer geometric focusing and electronic delay-andsum beamforming, as well as the double derivative scattering operator; and R(r) is a tissue reflectivity function expressing the fractional change in bulk modulus K(r) relative to a mean bulk modulus K0: The current model assumes a spatially invariant hpe(r), and thus, I(r), allowing the use of fast Fourierbased convolution (Bamber and Dickinson 1980; Hergum et al. 2009); see the latter reference for a discussion of the assumption’s validity. The current work compensates attenuation by using time gain compensation; a more accurate solution would modify vpe(r) to include a material impulse response (Szabo 2004, ch. 4). METHODS The experimental methods corresponding to the above-described theory are now presented. After discussion of the data collection, the calculation of the two convolution terms in eqn (1), namely, the imager response I(r) and the tissue response R(r), is described. The 2-D convolution was performed using a rapid third-party Fourier-based implementation (Luong 2009). The final step in generating the simulated US image consisted of calculating the envelope using the magnitude of the analytic function. All signal processing was carried out with MATLAB (Mathworks, Natick, MA, USA). Data collection A female French bulldog had a tumor (later found to be mastocytoma) in the inguinal region and underwent surgery to have it removed. The tumor was imaged pre- and post-operatively using a 47.0-mm-aperture, 3.3- to 10.0-MHz linear array (LA522 E, Esaote, Genoa, Italy) connected to an ULA-OP Research US
Volume 39, Number 10, 2013
system (MSD Lab, University of Florence, Florence, Italy). The system applied 9 dB/cm attenuation compensation and had access to pre-beamformed data, allowing the generation of US images with arbitrary pixel resolution and without dynamic range compression. The excised tissue was sent for further analysis to a pathologist, who stained sections of the samples with hematoxylin and eosin (H&E) and digitized it at 403 (116.25 nm/pixel) resolution and 24-bit color using a commercial digital slide scanner (Panoramic 250 Flash, 3-DHistech, Budapest, Hungary). The images were later downsampled to a resolution of 1860.00 nm/pixel. The local scientific ethics council (National Research Institute for Radiobiology and Radiohygiene, Hungary) approved the study, and the informed owner of the dog declared her consent to the study. Calculation of imager response I(r) As noted by Gao et al. (2009), the imager response may be approximated analytically, measured experimentally or simulated with a package such as Field II (Jensen and Svendsen 1992). In the current work, the last approach was taken, using the parameters provided by the manufacturers: 3 cycles at 4.7 MHz pulse-echo wavelet, 15.7-mm active aperture, 20-mm focus depth, 6-mm element height, 0.04-mm kerf, 0.254-mm pitch. Calculation of tissue response R(r) The tissue response is given by the fractional change in bulk modulus K. To estimate K from an H&E-stained histology image, the image was first decomposed into its hematoxylin and eosin components using multiplicative color deconvolution (Ruifrok and Johnston 2001). The image was then segmented into a set of possible components (cytoplasm, nucleus, fat, red blood cell) according to empirically chosen value ranges for the two components (Table 1). The segmentation procedure was also able to identify ground substance and collagen, but the level of these components in the examples studied is negligible.
Table 1. Values of bulk modulus K used to calculate the tissue response R(r) and ranges of normalized H&E levels used to segment the H&E histology images Tissue type
Bulk modulus (GPa)
Cytoplasm
2.19 (Doyle et al. 2009)
Nucleus Fat Red blood cell
3.26 (Baddour et al. 2005) 1.90 (Duck 1990) 2.44 (Urick 1947)
H&E 5 hematoxylin and eosin.
H–E range 0.10 , H , 0.18, E . 0.30 H . 0.18 H , 0.18, E , 0.20 Everything else
Histology-based simulations of US imaging d M. GY€ONGY et al.
RESULTS AND DISCUSSION The canine mastocytoma tumor had a maximum diameter of 20 mm. We ensured that the pre-operative ultrasound image and the ultrasound image obtained immediately after tumor removal were aligned so as to reflect this dimension. Because of geometric limitations in histology processing, the histology slice was not taken at the maximum diameter, so exact correspondence between the imaging planes of histology and US was not attained. During surgery and on the post-operative US image, partial hemorrhaging around the tumor was observed. Two histology slices were obtained, one without and the other with signs of hemorrhaging. These were paired with the pre- and post-operative US images. Figure 1 illustrates a comparison between simulated and clinical US images in the case without hemorrhaging. The Kolmogorov-Smirnov (KS) method was used to
1927
compare the empirical distributions of the simulated and clinical pixel values. The number N of independent samples in the comparison was estimated from the ratio of the 4-mm2 comparison windows to the simulated 26-dB imager response area of 0.156 mm2, namely, 26. Hence, for a significance level of 0.01, the KS value needs to be less than 0.31 (Lindley and Scott 1995). This is indeed the case for the two regions: tumor (KS 5 0.27) and stroma (KS 5 0.12). Figure 2 illustrates a comparison between simulated and clinical US images in the case with hemorrhaging. Again, the KS values indicate good agreement between simulations and experiment (tumor core: 0.17, low bleeding: 0.14, high bleeding: 0.05). The observed hypoechogenicity from high levels of bleeding correlates with previous observations that the level of backscatter in blood reaches a peak at 15–20% hematocrit because of a limited level of spatial variability in acoustic properties at high hematocrits (Mo et al. 1994).
Fig. 1. Comparison of simulation and experiment for data without signs of hemorrhaging. Red signifies a tumor region, and green signifies surrounding stroma. Top left: Original histology image. Top center: Resulting simulated ultrasound (US) image. Top right: Pre-operative US image. Bottom left: Close-up views of the highlighted regions. Bottom middle: Envelope statistics of simulated US regions. Values in parentheses are the KS values expressing the degree of dissimilarity between simulations and experiment. Bottom right: Envelope statistics of clinical US regions.
1928
Ultrasound in Medicine and Biology
Volume 39, Number 10, 2013
Fig. 2. Comparison of simulation and experiment for data with signs of hemorrhaging. Red signifies part of the tumor core; green, a region of low bleeding; and blue, a region of high bleeding. Top left: Original histology image. Top center: Resulting simulated ultrasound (US) image. Top right: Post-operative US image. Bottom left: Close-up views of the highlighted regions. Bottom middle: Envelope statistics of simulated US regions. Values in parentheses are the KS values expressing the degree of dissimilarity between simulations and experiment. Bottom right: Envelope statistics of clinical US regions.
CONCLUSIONS In this technical note, a method of simulating US images from whole-slide H&E-stained histology is described. The method has the potential to further the understanding of how tissue microstructures lead to US images, as well as enable a prior assessment of the suitability of a certain imaging system for detecting a given pathology. However, the method needs to undergo proper validation, along with possible refinement, before meeting such goals. In particular, the correctness of the shift-invariant convolution assumption needs to be rigorously ensured by either infinite transmit or synthetic aperture focusing; the mapping from histology to bulk
modulus needs to be validated and refined as needed using acoustic microscopy data; lastly, to ensure the generality of the model, the use of 3-D histology images is needed to capture the intricate 3-D structure of animal tissue. Nevertheless, the preliminary statistical comparisons between simulated and clinical US images herein presented provide an encouraging basis for further work. Acknowledgments—Miklos Gy€ongy thanks Enrico Boni, Tamas Garay, Barna Laszlo, Nora Meggyeshazi, Gabor Tornai and Piero Tortoli for their extensive technical help and Christian Coviello, Sarolta Karpati, Michael Oelze, Tamas Roska and Andras Szasz for their technical comments and encouragement. The research was funded by the Hungarian Government through Grant TAMOP-4.2.1.B-11/2/KMR2011–0002.
Histology-based simulations of US imaging d M. GY€ONGY et al.
REFERENCES Baddour RE, Sherar MD, Hunt JW, Czarnota GJ, Kolios MC. Highfrequency ultrasound scattering from microspheres and single cells. J Acoust Soc Am 2005;117:934–943. Bamber J, Dickinson R. Ultrasonic B-scanning: A computer simulation. Phys Med Biol 1980;25:463–479. Catalano O, Nunziata A, Siani A. Fundamentals in oncologic ultrasound: Sonographic imaging and intervention in the cancer patient. Berlin/New York: Springer; 2009. Cobbold RSC. Foundations of biomedical ultrasound. New York: Oxford University Press; 2006. Daoud M, Lacefield J. Stochastic modeling of normal and tumor tissue microstructure for high-Frequency ultrasound imaging simulations. IEEE Trans Biomed Eng 2009;56:2806–2815. Dapore AJ, King MR, Harter J, Sarwate S, Oelze ML, Zagzebski JA, Do MN, Hall TJ, O’Brien WD. Analysis of human fibroadenomas using three-dimensional impedance maps. IEEE Trans Med Imaging 2011;30:1206–1213. Doyle TE, Tew AT, Warnick KH, Carruth BL. Simulation of elastic wave scattering in cells and tissues at the microscopic level. J Acoust Soc Am 2009;125:1751–1767. Duck FA. Physical properties of tissue. San Diego: Academic Press; 1990. Gao H, Choi HF, Claus P, Boonen S, Jaecques S, van Lenthe G, Van Der Perre G, Lauriks W, D’hooge J. A fast convolution-based methodology to simulate 2-D/3-D cardiac ultrasound images. IEEE Trans Ultrason Ferroelectr Freq Control 2009;56:404–409. Hergum T, Langeland S, Remme E, Torp H. Fast ultrasound imaging simulation in k-space. IEEE Trans Ultrason Ferroelectr Freq Control 2009;56:1159–1167. Hung C, Changchien C, Lu S, Eng H, Wang J, Lee C, Hsu C, Tung H. Sonographic features of hepatic adenomas with pathologic correlation. Abdom Imaging 2001;26:500–506. Jensen J. A model for the propagation and scattering of ultrasound in tissue. J Acoust Soc Am 1991;89:182–190.
1929
Jensen J, Svendsen N. Calculation of pressure fields from arbitrarily shaped, apodized, and excited ultrasound transducers. IEEE Trans Ultrason Ferroelectr Freq Control 1992;39:262–267. Lindley DV, Scott WF. New Cambridge statistical tables. London/New York: Cambridge University Press; 1995. Luong B. FFT-based convolution. MATLAB Central File Exchange, #24504. Natick, MA: Mathworks; 2009. Mari JM, Blu T, Matar OB, Unser M, Cachard C. A bulk modulus dependent linear model for acoustical imaging. J Acoust Soc Am 2009; 125:2413–2419. Mo L, Kuo I, Shung K, Ceresne L, Cobbold R. Ultrasound scattering from blood with hematocrits up to 100%. IEEE Trans Biomed Eng 1994;41:91–95. Oelze M, O’Brien W, Zachary J. High frequency quantitative ultrasound imaging of solid tumors in mice. Acoust Imaging 2007;28:301–306. Ruifrok AC, Johnston DA. Quantification of histochemical staining by color deconvolution. Anal Quant Cytol Histol 2001;23:291–299. Saha R, Kolios M. Effects of cell spatial organization and size distribution on ultrasound backscattering. IEEE Trans Ultrason Ferroelectr Freq Control 2011;58:2118–2131. Stepanishen PR. Pulsed transmit/receive response of ultrasonic piezoelectric transducers. J Acoust Soc Am 1981;69:1815–1827. Szabo T. Diagnostic ultrasound imaging: Inside out. San Diego: Academic Press; 2004. Tsui P-H, Liao Y-Y, Chang C-C, Kuo W-H, Chang K-J, Yes C-K. Classification of benign and malignant breast tumors by 2-D analysis based on contour description and scatterer characterization. IEEE Trans Med Imaging 2010;29:513–521. Urick R. A sound velocity method for determining the compressibility of finely divided substances. J Appl Phys 1947;18:983–987. Vlad R, Saha R, Alajez N, Ranieri S, Czarnota G, Kolios M. An increase in cellular size variance contributes to the increase in ultrasound backscatter during cell death. Ultrasound Med Biol 2010;36: 1546–1558.
http://dx.doi.org/10.1016/j.ultrasmedbio.2013.05.005
d
Technical Note HISTOLOGY-BASED SIMULATIONS OF ULTRASOUND IMAGING: METHODOLOGY y x RA SZALAI,z and IMRE KALL ONGY,* LAJOS BALOGH, KLA O* MIKLOS GY€
* Faculty of Information Technology, Pazmany Peter Catholic University, Budapest, Hungary; y National Research Institute for Radiobiology and Radiohygiene, Budapest, Hungary; z Department of Dermatology, Dermatooncology and Venerology, Semmelweis University, Budapest, Hungary; and x Institute of Experimental Medicine, Hungarian Academy of Sciences, Budapest, Hungary (Received 20 November 2012; revised 2 May 2013; in final form 11 May 2013)
Abstract—Simulations of ultrasound (US) images based on histology may shed light on the process by which microscopic tissue features translate to a US image and may enable predictions of feature detectability as a function of US system parameters. This technical note describes how whole-slide hematoxylin and eosin-stained histology images can be used to generate maps of fractional change in bulk modulus, whose convolution with the impulse response of the US system yields simulated US images. The method is illustrated by two canine mastocytoma histology images, one with and the other without signs of intra-operative hemorrhaging. Quantitative comparisons of the envelope statistics with corresponding clinical US images provide preliminary validation of the method. (E-mail: gyongy. [email protected]) Ó 2013 World Federation for Ultrasound in Medicine & Biology. Key Words: Simulation, Modeling, Histology, Envelope statistics, Tissue microstructure, Ultrasonic tissue characterization.
processes such as apoptosis (Vlad et al. 2010). By locating the nuclei in histology slides and calculating certain microstructural statistics from these, a stochastic model has been used to simulate B-mode images whose first-order statistics match those of the corresponding, experimentally obtained images (Daoud and Lacefield 2009). In an alternative, continuum-based approach, 3-D acoustic maps were generated from histology slides and were used to estimate the effective acoustic scatterer diameter of various tissue types (Dapore et al. 2011). Neither model is naturally extensible to modeling higher-order features such as interfaces between tissue types, because only aggregate descriptors are obtained from limited-view histology slices. The current technical note expands on the above work by directly simulating US images from whole-slide histology and comparing their envelope statistics with those of clinical US images. It is hoped that the method described could form the basis of an experimentally validated histology-based model of US image formation.
INTRODUCTION A sufficiently accurate model of ultrasound (US) image formation from microscopic images of tissue could help explain how millimeter-scale features of clinical US images arise from micrometer cellular structures. The model could also be used to predict the viability of an US diagnostic method on a given transducer before clinical testing. Examples of applications are B-mode images of malignant tumors that usually (Catalano et al. 2009), but not always (Hung et al. 2001), appear hypoechoic, and quantitative US, where parameters derived from backscattered US (e.g., envelope statistics) promise better differential diagnosis (Oelze et al. 2007; Tsui et al. 2010). Theoretical models of US image formation abound in the literature, either assuming discrete scatterers or continuously varying acoustic maps (Mari et al. 2009). The discrete scatterer approach has been used to model tissue scattering as arising from clusters of cells (Doyle et al. 2009; Saha and Kolios 2011), helping to explain observed changes in US backscatter resulting from
THEORY OF IMAGE FORMATION Address correspondence to: Miklos Gy€ongy, Faculty of Information Technology, Pazmany Peter Catholic University, Prater utca 50/a, Budapest H-1083, Hungary. E-mail: [email protected]
Using common assumptions (small fluctuations from the mean speed of sound c0 and mean density r0, 1925
1926
Ultrasound in Medicine and Biology
single scattering) found in other continuum scattering models (Cobbold 2006; Jensen 1991), Mari et al. (2009) reported that scattering can be expressed as a function of the longitudinal bulk modulus K only. With suitable rearrangement of terms and a transformation between time t and axial distance z 5 c0t/2, the image u(r) as a function of mapped location r can be expressed as the convolution of an imager response I(r) with a tissue response R(r): uðrÞ 5 IðrÞ RðrÞ
(1)
IðrÞ 5 vpe ðrÞ hpe ðrÞ
(2)
where vpe(r) is the spatial-domain pulse-echo wavelet that includes the electrical excitation, as well as transmit and receive impulse responses, of the transducer (Stepanishen 1981); hpe(r) is the pulse-echo spatial impulse response of the transducer that incorporates transducer geometric focusing and electronic delay-andsum beamforming, as well as the double derivative scattering operator; and R(r) is a tissue reflectivity function expressing the fractional change in bulk modulus K(r) relative to a mean bulk modulus K0: The current model assumes a spatially invariant hpe(r), and thus, I(r), allowing the use of fast Fourierbased convolution (Bamber and Dickinson 1980; Hergum et al. 2009); see the latter reference for a discussion of the assumption’s validity. The current work compensates attenuation by using time gain compensation; a more accurate solution would modify vpe(r) to include a material impulse response (Szabo 2004, ch. 4). METHODS The experimental methods corresponding to the above-described theory are now presented. After discussion of the data collection, the calculation of the two convolution terms in eqn (1), namely, the imager response I(r) and the tissue response R(r), is described. The 2-D convolution was performed using a rapid third-party Fourier-based implementation (Luong 2009). The final step in generating the simulated US image consisted of calculating the envelope using the magnitude of the analytic function. All signal processing was carried out with MATLAB (Mathworks, Natick, MA, USA). Data collection A female French bulldog had a tumor (later found to be mastocytoma) in the inguinal region and underwent surgery to have it removed. The tumor was imaged pre- and post-operatively using a 47.0-mm-aperture, 3.3- to 10.0-MHz linear array (LA522 E, Esaote, Genoa, Italy) connected to an ULA-OP Research US
Volume 39, Number 10, 2013
system (MSD Lab, University of Florence, Florence, Italy). The system applied 9 dB/cm attenuation compensation and had access to pre-beamformed data, allowing the generation of US images with arbitrary pixel resolution and without dynamic range compression. The excised tissue was sent for further analysis to a pathologist, who stained sections of the samples with hematoxylin and eosin (H&E) and digitized it at 403 (116.25 nm/pixel) resolution and 24-bit color using a commercial digital slide scanner (Panoramic 250 Flash, 3-DHistech, Budapest, Hungary). The images were later downsampled to a resolution of 1860.00 nm/pixel. The local scientific ethics council (National Research Institute for Radiobiology and Radiohygiene, Hungary) approved the study, and the informed owner of the dog declared her consent to the study. Calculation of imager response I(r) As noted by Gao et al. (2009), the imager response may be approximated analytically, measured experimentally or simulated with a package such as Field II (Jensen and Svendsen 1992). In the current work, the last approach was taken, using the parameters provided by the manufacturers: 3 cycles at 4.7 MHz pulse-echo wavelet, 15.7-mm active aperture, 20-mm focus depth, 6-mm element height, 0.04-mm kerf, 0.254-mm pitch. Calculation of tissue response R(r) The tissue response is given by the fractional change in bulk modulus K. To estimate K from an H&E-stained histology image, the image was first decomposed into its hematoxylin and eosin components using multiplicative color deconvolution (Ruifrok and Johnston 2001). The image was then segmented into a set of possible components (cytoplasm, nucleus, fat, red blood cell) according to empirically chosen value ranges for the two components (Table 1). The segmentation procedure was also able to identify ground substance and collagen, but the level of these components in the examples studied is negligible.
Table 1. Values of bulk modulus K used to calculate the tissue response R(r) and ranges of normalized H&E levels used to segment the H&E histology images Tissue type
Bulk modulus (GPa)
Cytoplasm
2.19 (Doyle et al. 2009)
Nucleus Fat Red blood cell
3.26 (Baddour et al. 2005) 1.90 (Duck 1990) 2.44 (Urick 1947)
H&E 5 hematoxylin and eosin.
H–E range 0.10 , H , 0.18, E . 0.30 H . 0.18 H , 0.18, E , 0.20 Everything else
Histology-based simulations of US imaging d M. GY€ONGY et al.
RESULTS AND DISCUSSION The canine mastocytoma tumor had a maximum diameter of 20 mm. We ensured that the pre-operative ultrasound image and the ultrasound image obtained immediately after tumor removal were aligned so as to reflect this dimension. Because of geometric limitations in histology processing, the histology slice was not taken at the maximum diameter, so exact correspondence between the imaging planes of histology and US was not attained. During surgery and on the post-operative US image, partial hemorrhaging around the tumor was observed. Two histology slices were obtained, one without and the other with signs of hemorrhaging. These were paired with the pre- and post-operative US images. Figure 1 illustrates a comparison between simulated and clinical US images in the case without hemorrhaging. The Kolmogorov-Smirnov (KS) method was used to
1927
compare the empirical distributions of the simulated and clinical pixel values. The number N of independent samples in the comparison was estimated from the ratio of the 4-mm2 comparison windows to the simulated 26-dB imager response area of 0.156 mm2, namely, 26. Hence, for a significance level of 0.01, the KS value needs to be less than 0.31 (Lindley and Scott 1995). This is indeed the case for the two regions: tumor (KS 5 0.27) and stroma (KS 5 0.12). Figure 2 illustrates a comparison between simulated and clinical US images in the case with hemorrhaging. Again, the KS values indicate good agreement between simulations and experiment (tumor core: 0.17, low bleeding: 0.14, high bleeding: 0.05). The observed hypoechogenicity from high levels of bleeding correlates with previous observations that the level of backscatter in blood reaches a peak at 15–20% hematocrit because of a limited level of spatial variability in acoustic properties at high hematocrits (Mo et al. 1994).
Fig. 1. Comparison of simulation and experiment for data without signs of hemorrhaging. Red signifies a tumor region, and green signifies surrounding stroma. Top left: Original histology image. Top center: Resulting simulated ultrasound (US) image. Top right: Pre-operative US image. Bottom left: Close-up views of the highlighted regions. Bottom middle: Envelope statistics of simulated US regions. Values in parentheses are the KS values expressing the degree of dissimilarity between simulations and experiment. Bottom right: Envelope statistics of clinical US regions.
1928
Ultrasound in Medicine and Biology
Volume 39, Number 10, 2013
Fig. 2. Comparison of simulation and experiment for data with signs of hemorrhaging. Red signifies part of the tumor core; green, a region of low bleeding; and blue, a region of high bleeding. Top left: Original histology image. Top center: Resulting simulated ultrasound (US) image. Top right: Post-operative US image. Bottom left: Close-up views of the highlighted regions. Bottom middle: Envelope statistics of simulated US regions. Values in parentheses are the KS values expressing the degree of dissimilarity between simulations and experiment. Bottom right: Envelope statistics of clinical US regions.
CONCLUSIONS In this technical note, a method of simulating US images from whole-slide H&E-stained histology is described. The method has the potential to further the understanding of how tissue microstructures lead to US images, as well as enable a prior assessment of the suitability of a certain imaging system for detecting a given pathology. However, the method needs to undergo proper validation, along with possible refinement, before meeting such goals. In particular, the correctness of the shift-invariant convolution assumption needs to be rigorously ensured by either infinite transmit or synthetic aperture focusing; the mapping from histology to bulk
modulus needs to be validated and refined as needed using acoustic microscopy data; lastly, to ensure the generality of the model, the use of 3-D histology images is needed to capture the intricate 3-D structure of animal tissue. Nevertheless, the preliminary statistical comparisons between simulated and clinical US images herein presented provide an encouraging basis for further work. Acknowledgments—Miklos Gy€ongy thanks Enrico Boni, Tamas Garay, Barna Laszlo, Nora Meggyeshazi, Gabor Tornai and Piero Tortoli for their extensive technical help and Christian Coviello, Sarolta Karpati, Michael Oelze, Tamas Roska and Andras Szasz for their technical comments and encouragement. The research was funded by the Hungarian Government through Grant TAMOP-4.2.1.B-11/2/KMR2011–0002.
Histology-based simulations of US imaging d M. GY€ONGY et al.
REFERENCES Baddour RE, Sherar MD, Hunt JW, Czarnota GJ, Kolios MC. Highfrequency ultrasound scattering from microspheres and single cells. J Acoust Soc Am 2005;117:934–943. Bamber J, Dickinson R. Ultrasonic B-scanning: A computer simulation. Phys Med Biol 1980;25:463–479. Catalano O, Nunziata A, Siani A. Fundamentals in oncologic ultrasound: Sonographic imaging and intervention in the cancer patient. Berlin/New York: Springer; 2009. Cobbold RSC. Foundations of biomedical ultrasound. New York: Oxford University Press; 2006. Daoud M, Lacefield J. Stochastic modeling of normal and tumor tissue microstructure for high-Frequency ultrasound imaging simulations. IEEE Trans Biomed Eng 2009;56:2806–2815. Dapore AJ, King MR, Harter J, Sarwate S, Oelze ML, Zagzebski JA, Do MN, Hall TJ, O’Brien WD. Analysis of human fibroadenomas using three-dimensional impedance maps. IEEE Trans Med Imaging 2011;30:1206–1213. Doyle TE, Tew AT, Warnick KH, Carruth BL. Simulation of elastic wave scattering in cells and tissues at the microscopic level. J Acoust Soc Am 2009;125:1751–1767. Duck FA. Physical properties of tissue. San Diego: Academic Press; 1990. Gao H, Choi HF, Claus P, Boonen S, Jaecques S, van Lenthe G, Van Der Perre G, Lauriks W, D’hooge J. A fast convolution-based methodology to simulate 2-D/3-D cardiac ultrasound images. IEEE Trans Ultrason Ferroelectr Freq Control 2009;56:404–409. Hergum T, Langeland S, Remme E, Torp H. Fast ultrasound imaging simulation in k-space. IEEE Trans Ultrason Ferroelectr Freq Control 2009;56:1159–1167. Hung C, Changchien C, Lu S, Eng H, Wang J, Lee C, Hsu C, Tung H. Sonographic features of hepatic adenomas with pathologic correlation. Abdom Imaging 2001;26:500–506. Jensen J. A model for the propagation and scattering of ultrasound in tissue. J Acoust Soc Am 1991;89:182–190.
1929
Jensen J, Svendsen N. Calculation of pressure fields from arbitrarily shaped, apodized, and excited ultrasound transducers. IEEE Trans Ultrason Ferroelectr Freq Control 1992;39:262–267. Lindley DV, Scott WF. New Cambridge statistical tables. London/New York: Cambridge University Press; 1995. Luong B. FFT-based convolution. MATLAB Central File Exchange, #24504. Natick, MA: Mathworks; 2009. Mari JM, Blu T, Matar OB, Unser M, Cachard C. A bulk modulus dependent linear model for acoustical imaging. J Acoust Soc Am 2009; 125:2413–2419. Mo L, Kuo I, Shung K, Ceresne L, Cobbold R. Ultrasound scattering from blood with hematocrits up to 100%. IEEE Trans Biomed Eng 1994;41:91–95. Oelze M, O’Brien W, Zachary J. High frequency quantitative ultrasound imaging of solid tumors in mice. Acoust Imaging 2007;28:301–306. Ruifrok AC, Johnston DA. Quantification of histochemical staining by color deconvolution. Anal Quant Cytol Histol 2001;23:291–299. Saha R, Kolios M. Effects of cell spatial organization and size distribution on ultrasound backscattering. IEEE Trans Ultrason Ferroelectr Freq Control 2011;58:2118–2131. Stepanishen PR. Pulsed transmit/receive response of ultrasonic piezoelectric transducers. J Acoust Soc Am 1981;69:1815–1827. Szabo T. Diagnostic ultrasound imaging: Inside out. San Diego: Academic Press; 2004. Tsui P-H, Liao Y-Y, Chang C-C, Kuo W-H, Chang K-J, Yes C-K. Classification of benign and malignant breast tumors by 2-D analysis based on contour description and scatterer characterization. IEEE Trans Med Imaging 2010;29:513–521. Urick R. A sound velocity method for determining the compressibility of finely divided substances. J Appl Phys 1947;18:983–987. Vlad R, Saha R, Alajez N, Ranieri S, Czarnota G, Kolios M. An increase in cellular size variance contributes to the increase in ultrasound backscatter during cell death. Ultrasound Med Biol 2010;36: 1546–1558.