Ultrasound in Med. & Biol., Vol. 40, No. 12, pp. 2841–2850, 2014Copyright � 2014 World Federation for Ultrasound in Medicine & Biology

Printed in the USA. All rights reserved0301-5629/$ - see front matter

/j.ultrasmedbio.2014.07.021

http://dx.doi.org/10.1016

d Original Contribution

REAL-TIME MONITORING OF HIGH-INTENSITY FOCUSED ULTRASOUNDTHERMALTHERAPY USING THE MANIFOLD LEARNING METHOD

PARISA RANGRAZ,* HAMID BEHNAM,y POOYA SOBHEBIDARI,z and JAHAN TAVAKKOLIx

*Department of Biomedical Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran; yBiomedicalEngineering Department, School of Electrical Engineering, Iran University of Science and Technology, Tehran, Iran;

zDepartment of Electrical and Computer Engineering, Ryerson University, Toronto, Ontario, Canada; and xDepartment ofPhysics, Ryerson University, Toronto, Ontario, Canada

(Received 17 November 2013; revised 27 July 2014; in final form 31 July 2014)

ABiomeUniversrbiau.

Abstract—High-intensity focused ultrasound (HIFU) induces thermal lesions by increasing the tissue temperaturein a tight focal region. The main ultrasound imaging techniques currently used to monitor HIFU treatment arestandard pulse-echo B-mode ultrasound imaging, ultrasound temperature estimation and elastography-basedmethods. The present study was carried out on ex vivo animal tissue samples, in which backscattered radiofre-quency (RF) signals were acquired in real time at time instances before, during and after HIFU treatment. Themanifold learning algorithm, a non-linear dimensionality reduction method, was applied to RF signals whichconstruct B-mode images to detect the HIFU-induced changes among the image frames obtained during HIFUtreatment. In this approach, the embedded non-linear information in the region of interest of sequential imagesis represented in a 2-D manifold with the Isomap algorithm, and each image is depicted as a point on the recon-structedmanifold. Four distinct regions are chosen in themanifold corresponding to the four phases of HIFU treat-ment (before HIFU treatment, during HIFU treatment, immediately after HIFU treatment and 10-min after HIFUtreatment). It was found that disorganization of the points is achieved by increasing the acoustic power, and if thethermal lesion has been formed, the regions of points related to pre- and post-HIFU significantly differ. Moreover,the manifold embedding was repeated on 2-Dmoving windows in RF data envelopes related to pre- and post-HIFUexposure data frames. It was concluded that if meanvalues of the points related to pre- and post-exposure frames inthe reconstructed manifold are estimated, and if the Euclidean distance between these two mean values iscalculated and the sliding window is moved and this procedure is repeated for the whole image, a new image basedon the Euclidean distance can be formed in which the HIFU thermal lesion is detectable. (E-mail: p.rangraz@srbiau.ac.ir) � 2014 World Federation for Ultrasound in Medicine & Biology.

Key Words: High-intensity focused ultrasound, Thermal lesion detection, Manifold learning, Isomap algorithm.

INTRODUCTION

The goal of high-intensity focused ultrasound (HIFU), asa therapeutic application of high-energy acoustic waves,is to thermally coagulate the tissue at the focal spot of atransducer by focusing acoustic energy on this region(Gray et al. 2011). The primary bio-effect of HIFU treat-ment in the tissue is thermal lesion formation, and its sec-ondary effect is cavitation (Zhong et al. 2007). Manystudies over past decades have reported the monitoringand detection of HIFU thermal lesions using different ap-proaches such as magnetic resonance imaging (Christakiset al. 2004; Hynynen 2011; Khokhlova et al. 2009), X-ray

ddress correspondence to: Parisa Rangraz, Department ofdical Engineering, Science and Research Branch, Islamic Azadsity, Hesarak, Tehran, Iran 1477893855. E-mail: p.rangraz@ac.ir

2841

imaging (Siegel et al. 2000) and ultrasound imaging(Tavakkoli and Sanghvi 2011; Zhong et al. 2007).

There remain a few major hurdles before wide-spread clinical translation of HIFU surgery. Theseinclude the capability to efficiently monitor lesionformation and real-time characterization of the effectiveregion of treatment (thermal lesion) (Gray et al. 2011).One of the most straightforward methods of HIFU lesiondetection is ultrasound B-mode imaging (Arefiev et al.1998; Vaezy et al. 2001a). Conventional B-modeultrasound images depict variations in a medium’sechogenicity. A hyper-echoic region at the HIFU focususually indicates the formation of a lesion (Vaezyet al. 2001b). This hyper-echoic region has often beenconsidered to be associated with coagulation necrosisand the appearance of bubbles produced by acousticcavitation or boiling at the HIFU focus. However, the

2842 Ultrasound in Medicine and Biology Volume 40, Number 12, 2014

biological and physical mechanisms of hyper-echo for-mation are not yet fully understood (Zhang et al.2009). Conventional B-mode ultrasound imaging isnot sufficiently accurate and specific for reliable detec-tion of HIFU-induced thermal lesions (Bailey et al.2003; Dalong and Ebbini 2010; Seip et al. 2003; TerHaar 2007; Zhang et al. 2009). Tissue temperature canbe non-invasively estimated by detecting and measuringchanges in the speed of sound as a function of tempera-ture (Anand et al. 2007; Seip and Ebbini 1995).However, such factors as non-linearity (Bamber andHill 1979; Bloch et al. 1998), thermal expansion,different intensity levels (Damianou et al. 1997) and var-iations in speed of sound for different types of tissue(Miller et al. 2002) have so far imposed serious limita-tions on non-invasive temperature measurement usingultrasound. Measurement of change in ultrasound atten-uation can be used to monitor HIFU therapies. Ophiret al. (1984) extensively discussed various methods ofattenuation estimation using radiofrequency (RF) back-scattered ultrasound signals. Transient characteristics oftissue integrated backscatter (IBS), attenuation coeffi-cient and bubble activity as time traces before, duringand after HIFU treatment have been investigated tothis end (Rahimian and Tavakkoli 2013; Zhang et al.2009; Zhong et al. 2007). Recent studies have reportedthe potential for using acoustic radiation forceimpulses to monitor the changes in tissue scattererdisplacement that may be related to formation of theHIFU thermal lesion (Anand et al. 2010; Dalong andEbbini 2010; Hawa et al. 2010; Hou et al. 2011; Zhanget al. 2008). An ultrasonic Nakagami imagingtechnique has been proposed for monitoring HIFUthermal ablation in real time and has been found ableto detect the HIFU-induced thermal lesions (Li et al.2010). A non-invasive ultrasound-based technique wasalso developed to detect HIFU-induced thermal lesionsbased on the mechanical, acoustic and statistical proper-ties of tissues in vitro (Rangraz et al. 2012). Severalimportant aspects of ultrasound monitoring of HIFUtreatments still are not thoroughly understood. Thehyper-echoic region in the ultrasound B-mode imagemay not be a reliable indicator of HIFU lesions andcannot provide the separate transient characteristics ofbubble activity or local tissue acoustic properties thatdetermine the ultrasound energy absorption and scat-tering characteristics. Studies on changes in acousticpropagation properties of HIFU-treated tissues areimportant in understanding the transient characteristicsof biophysical mechanisms involved in real-time moni-toring of HIFU treatment. To date, there have been onlya few studies on simultaneous investigation of the dy-namic changes in acoustic properties of thermal lesions

and bubble activity over the entire HIFU treatment pro-cedure (Zhang et al. 2009).

In this article, a new ultrasound-based method ofHIFU thermal lesion detection is described that is capableof monitoring the trend of changes before, during and afterexposure of the tissue to HIFU using the manifold learningalgorithm (Roweis and Saul 2000; Tenenbaum et al. 2000).

There has been longstanding interest in reducing thedimensionality of data in various signal and image pro-cessing applications, while preserving relevant informa-tion (Nascimento and Silva 2010). Manifold learning isa method of non-linear data dimensionality reduction(Roweis and Saul 2000; Tenenbaum et al. 2000). In twonotable biomedical applications, it has been shown thatimage manifolds are useful representation tools inmagnetic resonance imaging sequences (Zhang andPless 2005; Zhang et al. 2005) and in automaticdetection of end-diastole and end-systole from echocardi-ography images (Gifani et al. 2010).

In this work, the Isomap algorithm (Tenenbaumet al. 2000) is implemented on the RF echo data envelope,and the non-linear embedded information in the region ofinterest of sequential images is represented in a 2-Dmani-fold where each frame of the image is depicted as a pointon the reconstructed manifold. Differences between theresults before, during and after HIFU treatment havebeen observed. Four distinct regions are chosen in themanifold corresponding to the four phases of HIFU treat-ment: before HIFU, during HIFU, immediately afterHIFU and 10 min after HIFU.

If the thermal lesion has been formed, regions ofpoints related to pre- and post-HIFU are significantlydifferent. Manifold embedding was repeated on movingwindows in the RF data envelope related to pre- andpost-HIFU exposure to investigate the applicability ofthe method in detecting HIFU thermal lesions. It wasfound that it is possible to detect the HIFU lesions bycalculating the mean values of the points of the coordi-nates related to pre-exposure in the reconstructed mani-fold. Moreover, the mean values of the points of thecoordinates related to post-HIFU exposure data frameswere estimated, and then the Euclidean distance betweenthese two mean values was calculated. Finally, a new im-age based on the calculated Euclidean distance wasformed in which the HIFU-induced thermal lesion wasdetectable. It was also found that disorganization of thepoints was achieved by increasing the acoustic power,and when the thermal lesion was formed, the regions ofpoints related to pre- and post-HIFU became significantlydifferent. It was also found that increasing the acousticpower results in some fluctuations associated with thelow-frequency signal, which could be related to bubbleactivity in the focal region of the HIFU beam.

Fig. 1. Confocal arrangement of the imaging probe with thehigh-intensity focused ultrasound (HIFU) transducer.

HIFU therapy monitoring using Manifold Learning method d P. RANGRAZ et al. 2843

METHODS

Manifold learning and Isomap algorithmsManifold learning algorithms are based on the intu-

ition that the dimensionality of data sets might be artifi-cially high and each data sample can be described as afunction of only a few underlying parameters (Roweisand Saul 2000; Tenenbaum et al. 2000). Hence, thesealgorithms attempt to uncover intrinsic parameters tofind a low-dimensional representation of the data. In otherwords, the data points are assumed to be samples from alow-dimensional manifold that is embedded in a high-dimensional space. Some examples of manifold learningalgorithms are: isometric feature mapping (Isomap)(Tenenbaum et al. 2000), locally linear embeddings(Saul and Roweis 2003) and local tangent space align-ment (Zhang and Zha 2004).

In this study, the applicability of the Isomap algo-rithm, as one of the main algorithms introduced for themanifold learning and non-linear dimensionality reduc-tion, was investigated with respect to the detection ofHIFU thermal lesions. In the Isomap algorithm, geodesicdistances between points are extracted instead of simplytaking the Euclidean distance; thus, a geometric distancegraph is constructed and the embedding is obtained fromthe graph (Lim et al. 2003; Wu and Chan 2004).

In the first step, the Isomap algorithm determineswhich points are neighbors on the manifold M, basedon the distances dXði; jÞ between pairs of points i, j inthe input space X. Two simple methods are (i) to connecteach point to all points within a fixed radius and (ii) toconnect each point to all of its specific numbers of nearestneighbors. These neighborhood relations are representedas a weighted graph G over the data points, with edges ofweight dXði; jÞ between neighboring points. In the secondstep, it estimates the geodesic distances dMði; jÞ betweenall pairs of points on the manifold M by computing theirshortest path distances dGði; jÞ in the graph G.

The final step applies classic multidimensionalscaling to the matrix of graph distances DG 5 fdGði; jÞg,constructing an embedding of the data in a d-dimensionalEuclidean space Y that best preserves the manifold’s esti-mated intrinsic geometry. The coordinate vectors yi forpoints in Y are chosen to minimize the cost function Edefined as

E5 ktðDGÞ2tðDYÞkL2 (1)

where DY denotes the matrix of Euclidean distancesfdYði; jÞ5

��yi2yj��g and kAkL2 is the L2 matrix norm

given asffiffiffiffiffiffiffiffiffiffiffiffiffiffiP

i;j A2ij

q. The operator t converts distances to

inner products, which uniquely characterize the geometryof the data in an efficient optimization form. The globalminimum of eqn (1) is achieved by setting the coordinates

yi to the top d eigenvectors of the matrix tðDGÞ. As inprincipal component analysis or multidimensionalscaling, the true dimensionality of the data can beestimated from the decrease in error as the dimensionalityof Y is increased (Lim et al. 2003; Tenenbaum et al. 2000).

Ultrasound RF echo data acquisitionIn this study, a spherically concave single-element

HIFU transducer (Model 6699 A101; Imasonic, Voraysur1’Ognon, France) with a center frequency of1 MHz, F-number of 0.8 and aperture diameter of125 mm was used to generate HIFU beams and inducethermal lesions in the tissue. The lateral width and axiallength of the focal spot were measured in water to be 2and 8 mm at full width at half maximum (FWHM),respectively. An ultrasound imaging system (Sonix RP,Ultrasonix, Richmond, BC, Canada) and a confocal en-docavity convex array probe (EC9-5/10, Ultrasonix)with 192 elements, center frequency of 6.5 MHz andbandwidth of 3 MHz were used to acquire B-mode im-ages and RF backscattered data frames. The confocalarrangement of the imaging probe and the HIFU trans-ducer is illustrated in Figure 1. The electrical signaldriving the HIFU transducer was generated by anAFG3101 Arbitrary Function Generator (Tektronix,Beaverton, OR, USA) and amplified with an AG1012RF power amplifier (T&C Power Conversions, Roches-ter, NY, USA).

Figure 2 is a schematic diagram of the experimentalsetup. The experiments were carried out on fresh ex vivoporcine muscle tissue samples. The acquired data includeRF echo signals and B-mode images from HIFU experi-ments with different total acoustic powers of 10, 30, 50,70, 90 and 110 W.

Fig. 2. Schematic diagram of the experimental setup for image-guided high-intensity focused ultrasound (HIFU).RF 5 radiofrequency.

2844 Ultrasound in Medicine and Biology Volume 40, Number 12, 2014

The RF data frames were captured before, during,immediately after and 10 min after each HIFU exposure.First, the RF data were acquired from the normal tissue(pre-HIFU data set); then, a HIFU thermal lesion wasinduced and the RF data were acquired during and imme-diately after the treatment. During treatment, RF datawere captured when the HIFU transducer was momen-tarily off for 120 ms to avoid acoustic and electrical inter-ference between the HIFU transducer and the ultrasoundimaging probe. Finally, post-HIFU data set were collected10 min after the HIFU treatment from the same target re-gion in the tissue. Figure 3 illustrates the timing sequenceof HIFU exposures and the acquisition of RF data and B-mode images. Total HIFU treatment time was 10 s; RFdata were acquired until 20 s post-exposure and then at10 min post-exposure. The before, during, immediatelyafter and 10 min after exposure data were collected fordifferent acoustic powers, as mentioned above.

Throughout the experiment, the tissue was fixed in-side a tissue holder. Therefore, all data sets were acquiredfrom the same target region of the tissue.

Fig. 3. Timing sequence of high-intensity focused ultrasound (H

HIFU thermal lesionsTo process the collected RF data in all experiments,

the manifold learning algorithm was implemented andapplied to RF echo data obtained from the target tissue re-gion where the HIFU lesions are expected to be induced.The algorithm was then applied to the images formedfrom the whole tissue. Figure 4 illustrates the thermal le-sions induced at total acoustic powers ranging from 10 to110W in steps of 20W. Figure 4 shows tissue samples cutand folded open, with the HIFU lesions placed in its mid-dle part identified by arrows. At the total acoustic powerof 10 W, the lesion could not been identified. As illus-trated in Figure 4 the depth of lesions from the tissue sur-face was measured. The tissue samples used in this studywere acquired from a local slaughterhouse.

RESULTS

All RF data processing and image formation taskswere performed in the Cartesian system of coordinates.The acquired frames comprised 192 lines with 4680

IFU) exposure and radiofrequency echo data acquisition.

Fig. 4. Tissue slices revealing high-intensity focused ultrasound lesions in their middle parts at acoustic powers of 10, 30,50, 70, 90 and 110 W.

HIFU therapy monitoring using Manifold Learning method d P. RANGRAZ et al. 2845

samples per line, which represents 90.1 mm of tissuedepth. The data acquisition sampling frequency was40 MHz. The lateral distance between two adjacentscan lines was 0.3 mm. For each acoustic power, 81frames (including 16 frames pre-HIFU, 40 frames during

Fig. 5. Two-dimensional non-linear embedding of high-intensimuscle tissues using the Isomap algorithm (with k 5 20 neighjections of original images include the lines connecting the symment are depicted by different colors. Green5 before HIFU exp

after HIFU exposure, magenta 5 1

HIFU, 9 frames immediately post-HIFU and 16 frames10 min post-HIFU) were acquired.

In the first step, a 2-D moving window was imple-mented axially and laterally in the region of interest of600 pixels3 20 lines, centered at the inducedHIFU lesion.

ty focused ultrasound (HIFU) treatment in ex vivo porcinebors) at different acoustic powers. These non-linear pro-bols from consecutive frames. The four phases of treat-osure, red5 during HIFU exposure, blue5 immediately0 min after HIFU exposure.

Fig. 6. Reconstructed B-mode images related to data collected after high-intensity focused ultrasound exposure. Acousticpowers range from 10 to 110 W in steps of 20 W.

2846 Ultrasound in Medicine and Biology Volume 40, Number 12, 2014

A total of 81 frames were used in this study, which resultedin the dimension of 81 3 12,000 data points in the firstspace (D). In the second step, the images were embeddedin a 2-D space by applying the Isomap algorithm. This re-sulted in a dimension reduction to 81 3 2 on the recon-structed manifold. Figure 5 illustrates the depicted pointson the reconstructed manifolds of the HIFU treatments atdifferent powers. In this figure, each symbol is a projectionof the original consecutive images connected to each othervia a line. As illustrated in Figure 5, there are four separateregions (green, red, blue and magenta) corresponding todata collected before, during, immediately after and10 min after HIFU treatment, respectively. It is seen thatthe shapes of the manifolds differ at the different totalacoustic powers used in this study.

Moreover, Figure 5 illustrates that the Euclidean dis-tance between the points pre- and post-HIFU increaseswhen acoustic power increases.

The Isomap algorithm was also implemented on thewhole RF data set acquired from each tissue sample pre-and post-HIFU exposure to detect the HIFU lesion. Thesize of the 2-D moving window in the Isomap algorithmwas 200 pixels3 4 lines, moving 30 pixels axially and 2lines laterally. This window size was chosen based on atrade-off between the region of interest in the tissueconsidered to obtain a local measurement of changes intissue and an acceptable variance of the estimation. Themean value of embedded points on the reconstructedmanifold from each window related to the 16 frames ofpre-HIFU was calculated. This value was then estimatedamong the embedded points of each window related to 16

frames of post-HIFU data and the Euclidean distance be-tween them that was estimated and plotted as an image. InFigure 6 are the constructed post-HIFU B-mode imagesof all treatments, which were chosen for implementationof the new algorithm. Figure 7 (a) illustrates the results ofthe Isomap algorithm for all treatments. Figure 7 (b) illus-trates the results when the surfaces of tissue samples wereremoved from data to improve thermal lesion detection.The color bar in Figure 7 represents the distance betweenthe mean values of coordinates of points pre- and post-HIFU treatment on the reconstructed manifold from theIsomap algorithm. Figure 7 can be compared with the le-sions identified in Figure 4. By increasing the acoustic po-wer, the lesions are clearer in both Figures 4 and 7 (b).One should note that the images in Figure 7 are in the Car-tesian system of coordinates. As a notable result, thedepth of the lesion in Figure 7 (b) at a total acoustic powerof 90 W was estimated to be about 1.8 cm. The measuredlesion depth in the experimental results obtained at thispower, illustrated in Figure 4, was 2.1 cm.

Figure 8 illustrates the values of each dimension ofdepicted points on the reconstructed manifold fromFigure 5 plotted as separate curves (red and blue) forthe 81 frames studied. It is seen that by increasing theacoustic power, the trends of the two curves becomeincreasingly different from each other.

DISCUSSION

The creation of HIFU thermal lesions is a highlynon-linear process; thus, linear dimension reduction

Fig. 7. Images generated on the basis of the Euclidean distance between points before and after high-intensity focusedultrasound treatment in the reconstructed manifold using the Isomap algorithm. (a) Results of Isomap algorithm for alltreatments. (b) Results of Isomap algorithm when the surfaces of tissues were removed from the region of interest.

HIFU therapy monitoring using Manifold Learning method d P. RANGRAZ et al. 2847

Fig. 8. Value of each dimension of depicted points on the reconstructed manifold (illustrated in Fig. 5) plotted as separatecurves (red and blue) for 81 frames.

2848 Ultrasound in Medicine and Biology Volume 40, Number 12, 2014

methods cannot recognize the correct relationships be-tween the RF data frames obtained during a typicalHIFU treatment. The Isomap algorithm is one of themost widely used manifold learning and non-linear datadimensionality reduction methods. The results of the pro-posed method indicate its potential in analyzing RF echodata sets acquired during a typical HIFU treatment. AsFigure 5 illustrates, the depicted points in two dimensionsfollowed a closed loop when there was no lesion. On theother hand, formation of the thermal lesion and the in-crease in acoustic power caused disorganization of thetrend of consecutive frames, and the points correspondingto pre- and post-HIFU treatment become separated.Figure 7 illustrates that by use of the Euclidean distance,the HIFU thermal lesion can be detected; but the resultsrelated to the surface of the tissue are different from thosefor other parts of the tissue and cause false detection ofthe thermal lesion by the Euclidean distance. Therefore,removing the tissue surface from the region of interestsignificantly improves lesion detection.

Recently, Rangraz et al. (2014) estimated dynamicchanges in the attenuation coefficient, scaling parameterof Nakagami distribution and tissue vibration duringHIFU treatment to detect thermal lesions ex vivo. Thevalues of dynamic changes were estimated by (i) spatiallyaveraging the attenuation coefficient and scaling param-eter of Nakagami distribution and (ii) estimating tissuevibration, both axially and laterally, in a region of interest83 103 8mm in the tissue. During HIFU exposure thereare structural changes in the tissue that affect the esti-mated parameters. The main effect of HIFU treatment

in tissue is due to thermal coagulation of structural pro-teins, a process known as protein denaturation. Structuralproteins are proteins that form the extracellular matrixand the cytoskeleton of all cells. The changes in tissuestructure caused by thermal coagulation are due mainlyto breakdown of structural proteins and cell shrinkagein the tissue (Thomsen 1999). Bubble activity (eithercavitation or boiling bubbles) present in the focal regionduring a typical HIFU exposure also affect the results.But estimating the standard deviations of parametersalong frames (including values before, during and afterexposure) could lead to estimation of the changes causedonly by tissue structural effects and detection of theinduced thermal lesion (Rangraz et al. 2014). In the pre-sent study, a 2-D moving window was implementedaxially and laterally in the region of interest (600pixels 3 20 lines) centered at the induced HIFU lesion.The images were embedded in a 2-D space by applyingthe Isomap algorithm, resulting in a dimension reductionon the reconstructed manifold. In addition, the meanvalues of coordinates of embedded points on the recon-structed manifold from each window related to the 16frames of pre-HIFU data were calculated. This valuewas then estimated among the embedded points of eachwindow related to 16 frames of the post-HIFU data.The Euclidean distance between them was subsequentlyestimated and plotted as an image. Compared with ourprevious works, in the present study we reduced thedimension of the data set and then studied the trend inchanges related to both structural and bubble effects,with the aim of separating these effects in a future study.

HIFU therapy monitoring using Manifold Learning method d P. RANGRAZ et al. 2849

With the proposed method, thermal lesion formation canbe detected when there is a clear distance between the co-ordinates of the embedded points pre and post-exposure.As illustrated in Figure 7, we estimated the distance be-tween mean values of the coordinates of embedded pointson the reconstructed manifold from each window relatedto the 16 frames of pre-HIFU and post-HIFU data.

CONCLUSIONS

In this study, we experimentally illustrated the feasi-bility of monitoring HIFU thermal ablation in ex vivo tis-sue by using the manifold learning algorithm. Theproposed HIFU lesion imaging method is based on esti-mating Euclidean distance between the mean values ofthe coordinates of points of pre- and post-HIFU treat-ment. The reconstructed manifold from the Isomap algo-rithm could detect thermal lesions in the cases used in thisstudy even when there was no significant bubble forma-tion activity in the focal region, which was needed inB-mode imaging monitoring (Zhang et al. 2009).Because the complexity of the manifold learning algo-rithm is relatively low, it can readily be integrated intopost-processing in current ultrasound imaging systemsfor real-time visualization of HIFU thermal lesions. Todevelop the method, the trend of values in each dimensionof depicted points on the reconstructed manifold of RFdata frames, which is shown in Figure 8, may be modeledby appropriate equations to separate characteristics ofthermal effects and bubble activity. This means that atlow acoustic power (e.g., 30W) the trend can be represen-tative of thermal effects, and by increasing the acousticpower, the trend becomes representative of both thermaleffects and bubble activity.

Acknowledgments—We thank Arthur Worthington from the Departmentof Physics, Ryerson University, Toronto, Canada, for his help in exper-imental setup and data acquisition. This work was partially supported byan Ontario Research Fund Research Excellence grant and a Natural Sci-ences and Engineering Research Council of Canada Discovery grantawarded to J. Tavakkoli.

REFERENCES

Anand A, Petruzzello J, Zhou S, Sethuraman S, Azevedo J. Two-dimensional real-time ultrasound technique to control lesion sizeduring HIFU therapy. In: Proceedings, 9th International Symposiumon Therapeutic Ultrasound, ISTU 2010. AIP Conf Proc2010;1215:79–82.

Anand A, Savery D, Hall C. Three-dimensional spatial and temporaltemperature imaging in gel phantoms using backscattered ultra-sound. IEEE Trans Ultrasound Ferroelectr Freq Control 2007;54:23–31.

Arefiev A, Prat F, Chapelon JY, Tavakkoli J, Cathignol D. Ultrasound-induced tissue ablation: Studies on isolated perfused porcine liver.Ultrasound Med Biol 1998;24:1033–1043.

Bailey MR, Khokhlova VA, Sapozhnikov OA, Kargl SG, Crum LA.Physical mechanisms of the therapeutic effect of ultrasound (a re-view). Acoust Phys 2003;49:369–388.

Bamber JC, Hill CR. Ultrasonic attenuation and propagation speed inmammalian tissues as a function of temperature. Ultrasound MedBiol 1979;5:149–157.

Bloch S, Michael R, Crum LA, Kaczkowski PJ, Keilman GW,Mourad PD. Measurements of sound speed in excised tissue overtemperatures expected under high-intensity focused ultrasound con-ditions. J Acoust Soc Am 1998;103:2868.

Christakis D, Pavlou M, Velev O, Kyriakou K, Trimikliniotis M. Highintensity focused ultrasound ablation of kidney guided by MRI. Ul-trasound Med Biol 2004;30:397–404.

Dalong L, Ebbini E. Real-time 2-D imaging of thermal and mechanicaltissue response to focused ultrasound. In: Proceedings, 9th Interna-tional Symposium on Therapeutic Ultrasound, ISTU 2009. AIPConf Proc 2010;1215:57–61.

Damianou CA, Sanghavi NT, Fry FJ, Maass-Moreno R. Dependence ofultrasonic attenuation and absorption in dog soft tissue on tempera-ture and thermal dose. J Acoust Soc Am 1997;102:628–634.

Gifani P, Behnam H, Shalbaf A, Alizadeh Sani Z. Automatic detectionof end-diastole and end-systole from echocardiography images us-ing manifold learning. Physiol Meas 2010;31:1–13.

Gray Y, Jianwen L, Fabrice M, Caroline M. Performance assessment ofHIFU lesion detection by harmonic motion imaging for focused(HMIFU): A 3-D finite-element-based framework with experi-mental validation. Ultrasound in Med Biol 2011;37:2013–2027.

Hawa C, Arora M, Coussious C, Noble A. The effect of attenuation co-efficient on radiation force impulse monitoring of thermal lesions.In: Proceedings, 9th International Symposium on Therapeutic Ultra-sound,ISTU 2010. AIP Conf Proc 1215:66–73.

Hou G, Lou J, Marquet F, Vappou J, Konofagou E. Performance assess-ment of HIFU lesion detection by harmonic motion imaging forfocused ultrasound (HMIFU): A 3-D finite-element-based frame-work with experimental validation. Ultrasound Med Biol 2011;37:2013–2027.

Hynynen K. MRIgHIFU: A tool for image-guided therapeutics. J MagnReson Imaging 2011;34:482–493.

Khokhlova TD, Canney MS, Lee D, Marro K, Crum LA, Khokhlova V,et al. Magnetic resonance imaging of boiling induced by high inten-sity focused ultrasound. J Acoust Soc Am 2009;125:2420–2431.

Li ML, Li DW, Liu HL, Lin MS. Ultrasonic Nakagami visualization ofHIFU-induced thermal lesions. In: Proceedings, IEEE InternationalUltrasonics Symposium (IUS) IEEE, San Diego, California, USA,11–14 October 2010. New York: IEEE; 2010. p. 2251–2253.

Lim SI, de Heras Ciechomski P, Sarni S, Thalman D. Planar arrange-ment of high-dimensional biomedical data sets by Isomap coordi-nates. In: Proceedings, 16th IEEE Symposium on Computer-BasedMedical Systems, CBMS 2003. New York: IEEE; 2003. p. 50–55.

Miller NR, Bamber JC, Meany PM. Fundamental limitations of non-invasive temperature imaging by means of ultrasound echo strainestimation. Ultrasound Med Biol 2002;28:1319–1333.

Nascimento JC, Silva JG. Manifold Learning for Object Tracking withMultiple Motion Dynamics. ECCV 2010, part III, LNCS 6313,2010; 172-185.

Ophir J, Shawker T, Makland NF, Miller JG. Attenuation estimation inreflection: Progress and prospects. Ultrason Imaging 1984;6:349–395.

Rahimian S, Tavakkoli J. Estimating dynamic changes of tissue attenu-ation coefficient during high intensity focused ultrasound treatment.J Ther Ultrasound 2013;1:14.

Rangraz P, BehnamH, ShakhssalimN, Tavakkoli J. A feed-forward neu-ral network algorithm to detect thermal lesions induced by high in-tensity focused ultrasound in tissue. J Med Signals Sensors 2012;2:16–32.

Rangraz P, Behnam H, Tavakkoli J. Nakagami imaging for detectingthermal lesions induced by high intensity focused ultrasound in tis-sue. Proc Inst Mech Eng H 2014;228:19–26.

Roweis ST, Saul LK. Nonlinear dimensionality reduction by locallylinear embedding. Science 2000;290:2323–2326.

Saul L, Roweis S. Think globally, fit locally: Unsupervised learning oflow dimensional manifolds. Mach Learn Res 2003;4:119–155.

Seip R, Ebbini E. Noninvasive estimation of tissue temperature responseto heating fields using diagnostic ultrasound. IEEE Trans BiomedEng 1995;42:828–839.

2850 Ultrasound in Medicine and Biology Volume 40, Number 12, 2014

Seip R, Tavakkoli J, Wunderlich A, Narendra TS, Dines K, CrumL. Real-time detection of multiple lesions during high intensity focusedultrasound (HIFU) treatments. In: Proceedings, 2nd InternationalSymposium on Therapeutic Ultrasound, Seattle, Washington, USA,29 July–1August 2002. Seattle: Center for Industrial &Medical Ultra-sound, Applied Physics Laboratory, University of Washington, 2003.

Siegel RJ, Atar S, Fishbein MC, Brasch AV, Peterson TM, et al. Nonin-vasive, transthoracic, low-frequency ultrasound augments thrombol-ysis in a canine model of acute myocardial infarction. Circulation2000;101:2026–2029.

Tavakkoli J, Sanghvi NT. Ultrasound-guided HIFU and thermal abla-tion. In: Frenkel V, (ed). Therapeutic ultrasound: Mechanisms toapplications, Nova Science Publishers. Hauppauge, NY: Nova Sci-ence; 2011. p. 137–161.

Tenenbaum JB, Silva V, Langford J. A global geometric framework fornon-linear dimensionality reduction. Science 2000;290:2319–2323.

Ter Haar G. Therapeutic applications of ultrasound. Prog Biophys MolBiol 2007;93:111–129.

Thomsen S. Mapping of thermal injury in biologic tissues using quanti-tative pathologic techniques. Proc SPIE 1999;3594:82–95.

Vaezy S, AndrewM,Kaczkowski P, CrumL. Imageguided acoustic ther-apy. Annu Rev Biomed Eng 2001a;3:375–390.

Vaezy S, Shi X,Martin RW, Chi E, Nelson PI, BailyMR, CrumLA. Real-time visualization of high-intensity focused ultrasound treatment us-ing ultrasound imaging. Ultrasound Med Biol 2001b;27:33–42.

Wu Y, Chan K. An extended Isomap algorithm for learning multi-classmanifold. In: Proceedings, 2004 International Conference on Ma-chine Learning and Cybernetics. Shanghai, China, August2004;6:3429–3433.

Zhang M, Castaneda B, Christensen J, Saad W, Bylund K, Hoyt K,Strang JG, Rubens DJ, Parker KJ. Real-time sonoelastography of he-patic thermal lesions in a swine model. Med Phys 2008;35:4132–4141.

Zhang Q, Pless R. Segmenting cardiopulmonary images using manifoldlearning with level sets. In: Computer vision for biomedical imageapplications. Berlin/Heidelberg: Springer; 2005. p. 479–488.

Zhang Q, Souvenir R, Pless R. Segmentation informed by manifoldlearning. In: Rangarajan A, Vemuri B, Yuille AL, (eds). Energyminimization methods in computer vision and pattern recognition.Berlin/Heidelberg: Springer; 2005. p. 398–413.

Zhang S, Wan M, Zhong H, Xu C, Liao Z, Liu H, Wang S. Dynamicchanges of integrated backscatter, attenuation coefficient and bubbleactivities during high intensity focused ultrasound (HIFU) treat-ment. Ultrasound Med Biol 2009;35:1828–1844.

Zhong H, Wan M, Jiang Y, Wang S. Monitoring imaging of lesionsinduced by high intensity focused ultrasound based on differentialultrasonic attenuation and integrated backscatter estimation. Ultra-sound Med Biol 2007;33:82–94.

Zhang Z, Zha H. Principal manifolds and non-linear dimensionalityreduction via tangent space alignment. SIAM J Sci Comput 2004;26:313–338.