Printed in the USA. All rights reserved0301-5629/08/$–see front matter
doi:10.1016/j.ultrasmedbio.2007.08.003
● Original Contribution
ULTRASOUND OF THE FINGERS FOR HUMAN IDENTIFICATIONUSING BIOMETRICS
GANESH NARAYANASAMY,*† J. BRIAN FOWLKES,*† OLIVER D. KRIPFGANS,*†
JON A. JACOBSON,* MICHEL DE MAESENEER,* RAINER M. SCHMITT,‡ andPAUL L. CARSON*†
*Department of Radiology and †Applied Physics Program, University of Michigan, Ann Arbor, MI, USA; and‡Cross Match Technologies Inc., Palm Beach Gardens, FL, USA
(Received 26 March 2007; revised 29 June 2007; in final form 2 August 2007)
Identification of individuals might become important inevery society for safety and security of its people andtheir assets. Current identification methods use finger-print systems that are susceptible to duplication andmismatch (Pankanti et al. 2001). Also, fingerprints do notindicate the vital and physiological status of the individ-ual, i.e., whether the person is alive or dead or extremelyagitated. Even with the use of new and emerging tech-nologies, human identification has been a challenge dueto concerns of accuracy, usability, privacy and securityof individuals involved (Jain et al. 1997, 2004). Theadvantages of correct identification include safer medicalcare as well as safer societies, reduced fraud and user-friendly man-machine interfaces (Jain et al. 2004).
Address correspondence to: Ganesh Narayanasamy, 200 Zina
Pitcher Place, 3316 Kresge-III, Ann Arbor, MI 48109, USA. E-mail:[email protected]
392
Here, we demonstrate the utility of ultrasound (US)in identifying potential biometrics that could supplementexisting identification methods. If successful, some ofthe measures can be achieved with inexpensive, newimaging systems. The internal anatomical structure ofhuman fingers from gray-scale US imaging offers a num-ber of such identifiers. These metrics can be evaluated bytrained observers or, possibly, in the future by automatedsegmentation of image features and analysis thereof.Such techniques are investigated extensively in medicalimaging as computer sided diagnosis (CAD) (Giger et al.1999; Sahiner et al. 2004) or tissue characterization(Golub et al. 1993). A third, newer approach is to matchthe entire image or image volume to those of knownindividuals by maximizing a similarity measure(s) andsetting threshold(s) for the degree of matching. Thisapproach is often used in two-dimensional (2D) im-ages in the fields of machine vision and human facerecognition (Neemuchwala et al. 2006) and has more
recently been applied in the volumetric image sets
Ultrasound of the fingers for human identification ● G. NARAYANASAMY et al. 393
acquired in medical imaging (Meyer et al. 1997; Parket al. 2004).
These same and related techniques can be applied tocolor flow Doppler images of the finger vasculature,although the vessel flow is much more variable over timefor a given individual, than are the structural shapes andsizes seen in gray-scale imaging. Peripheral perfusionmay be increased with exercise to allow more consistentimaging of a substantial portion of the vascular tree.Similarly, abnormally high perfusion or heart-rate mightbe employed as an indicator of individuals who are underhigh stress and perhaps not acting of their own freeaccord in conducting a legal or financial transaction.
MATERIALS AND METHODS
Studies were performed on individuals of bothgenders, hand dominance, various races or ethnicitiesand age groups. A total of 20 volunteers were studiedwith mean age of 33 � 16 y, mean height of 5 feet 7inches � 4 inches and mean weight of 146.5 � 24pounds. We imaged the distal portions of the three fin-gers—index, middle and ring—from both hands of allthe subjects both before and after a physical exerciseroutine. As a result of technical complications and timeconstraints, 36 out of 40 pairs of images were collected.Due to computational overhead, 15 matched pairs and 12unmatched pairs were randomly selected for image reg-istration study from among the before and after exerciseimage pairs collected. For the human reader study, all ofthe 36 pairs were used. A set of five subjects werescanned separately for reader training purposes and datafrom this set was not included in the analysis. The entirestudy was reviewed and approved by our InstitutionalReview Board (IRB) and informed consent was obtainedfrom each participant.
Each hand of all volunteers was scanned usinggray-scale, power Doppler and time permitting, com-pound modes (Entrekin et al. 1999). The scanning appa-ratus consisted of automated motorized two-axis transla-tion stage to move the transducer holder as shown inFig. 1. This apparatus sits on top of a TPX (4-methylpen-tene-1 based polyolefin) plate of 2.5 mm thickness thatseparates it from the finger. The volunteer was seatedcomfortably to minimize any motion artifacts with his orher fingers pressed lightly against the paddle after apply-ing US gel for good coupling. The scans were performedby translating the probe in the elevational directionacross the fingers by acquiring 2D images at a 0.4 mminterval. Translation speed was 2 mm/s and the USsystem trigger rate was set at 5 Hz to assure completionof each frame. The three-dimensional (3D) image sizewas about 500 � 200 � 125 pixels. The sonographic
equipment used in this study was a commercially avail-
able GE Logiq-9 scanner (General Electric Medical Sys-tems, Milwaukee, WI, USA), modified to fire on inputtrigger pulses.
The GE M12L array (General Electric Medical Sys-tems) was used with a transmit carrier frequency of 14MHz for standard b-mode gray-scale and compoundimaging with seven angles (0°, �6°, �9° and �12°). ForDoppler in color flow imaging, the transmit frequencywas 6.6 MHz and the pulse repetition frequency and wallfilter settings were 800 and 45 Hz or 400 and 26 Hz,respectively, depending on flash artifacts. The Dopplerstudy was cardiac-gated to allow acquisition of eachimage frame at the peak local systole (Bhatti et al. 2001).With minimum color flow signal averaging, this gavepeak color flow signal with no degradation of colorspatial resolution in the elevation direction. The neces-sary triggering was provided by taking the waveformoutput from an electrocardiogram (ECG) monitor(DINAMAP PRO 1000®, General Electric Medical Sys-tems) via a 3.5 mm socket and using an R-wave (theinitial upward deflection of the QRS complex that rep-resents electrical activity due to the depolarization of theventricles) threshold detector circuit with adjustable trig-ger level to produce a TTL pulse after the arrival of theR-wave. A time delay of 250 ms after the R-wave wasused before capturing a 2D Doppler image. The trans-ducer is then translated to its new location and anothertime delay of 50 ms is then implemented to stabilize the
Fig. 1. Experimental set-up: apparatus to scan the finger and theTPX plate with the US scanner in the background. X and Yaxes stepping motors for the transducer holders were on aframe attached to an upper compression paddle that helpsstabilize the finger. Shown here, but not used in this study, are
the vertical sliders and the bottom compression paddle.
scanning assembly and reduce the motion artifacts in the
394 Ultrasound in Medicine and Biology Volume 34, Number 3, 2008
Doppler image. This process gets repeated for the lengthof the scan across the three fingers.
Semiautomatic identificationRegistration software (described below) was used to
align any two 3D image sets using 3D transformations(translation, rotation and scale) and the classical Shannonmutual information (MI) metric. This alignment shouldcorrect for any distortion due to repositioning of fingers.A minimal transform is required, given that the subjectfingers were placed in the scanning apparatus in a similarorientation each time, leading primarily to a simple trans-lation transformation. Therefore, as will be seen later, themagnitude of rotational and scaling transformationsneeded in an attempt to align volume data sets provideda measure of match.
Fig. 2. Two-dimensional composite image of the central por-tion of the middle finger shown in a 3 � 8 checkerboard patternof rectangular blocks, alternatively from the reference imageand its registered homologous image, marked on respectiveblocks as R and H. The continuity of the finger structures atdifferent echo signal levels in some areas can be seen. Note thatthe bottom and right hand side of the image has been cropped
as they extended beyond the region-of-interest.
image of central portion of middle finger is shown aschecker-board (3 � 8) pattern of alternating referenceimage and its registered homologous image (marked as Rand H, respectively).
Visual matching of images in reader studyFor the reader study, a group of individuals that
included two expert musculoskeletal radiologists andtwo relative novices (physicists with backgrounds inultrasound imaging who were trained for this study bythe senior musculoskeletal radiologist) were selected.Readers performed 1-to-1 matching of pairs of 3D datasets using visualization and analysis software, ImageJ,developed at the National Institute of Health (Rasband1997 to 2007). The readers went through the pair ofstacks of 2D images to locate the corresponding struc-tures. Matching of data sets was based on certain criteriadefined by the senior radiologist and numerical scoreswere assigned according to a grading scale (describedbelow). Some of the finger structures that were chosen bythe senior radiologist for matching included the bonecontour, volar plate thickness at the distal interphalan-geal joint, flexor tendon thickness, curvature and inser-tion point onto the bone, length of the distal phalangeand the overall tissue appearance including regions withdistinctive echogenicity and shadowing (see Fig. 3). Inthe training session (where separate sets of data wereused that had no bearing on the final outcome), five pairsof 3D US images were given to each reader individuallywho was asked to record similarities or dissimilarities inthe finger anatomy. For an identical match, the structuresin all three fingers have to lie in the same location in thetwo image volumes.
Once trained, readers performed the same matchingprocedure on all 36 pairs of data sets. These data setswere divided into two groups; 18 pairs from the sameindividual and another 18 pairs from different individu-als. Each reader was offered the 36 pairs in a randomizedfashion without the reader knowing that there were anequal number of matched pairs and mismatched pairs.The readers reviewed the image pairs across two or three
Fig. 3. Primary internal finger structures employed in the visualmatching study.
sessions to limit fatigue. The results were recorded based
Ultrasound of the fingers for human identification ● G. NARAYANASAMY et al. 395
on a 1 to 5 grading scale (5 � highly likely match; 4 �likely match; 3 � inconclusive; 2 � unlikely match; 1 �highly unlikely match).
Cardiac-gated Doppler studyThis study was performed to determine the possi-
bility of using the US to evaluate the vascular patternsand record the variation in the blood flow with physicalexercise as a surrogate for stress level changes resultingin the blood flow modification. High levels of stressindicated by increase in the Doppler detected blood vol-ume could mean the individual is not acting out of his orher own free accord. Potentially, the power Doppler dataon finger could be used as a biometric identifier bymatching the vascular pattern.
We examined the variation of detected blood vol-ume, using color pixel density (CPD), with exerciseusing power Doppler US imaging of finger. The subjectperformed an exercise routine (jogging in place for 1 to2 min) that elevates the heart-rate by approximately 50%above the resting heart-rate. This exercise was chosen ata low risk level to raise the heart-rate of the subjects sothat the peripheral blood flow would increase among alarge fraction of subjects studied. After elevation of theheart-rate by approximately 50%, there was a decrease inthe heart-rate before the scan (approximately 1 min timedelay in recoupling the subject for the US scan) andduring the scan (2 min duration). Thirty-one Dopplerscan image pairs were collected from either hand of the20 individuals scanned due to technical difficulties andtime constraints.
Using US to image the vascular structure inside thefinger gave another set of structures as a potential bio-metric for matching. In this study, we imaged the vas-cular pattern using power Doppler and measured variousblood flow related quantities approximately (Carsonet al. 1993; Rubin et al. 1995). This should not onlyindicate the vital status of an individual but also theperson’s physiological status including anxiety level.Unlike the previous potential metrics, the blood flowvaries substantially within an individual from time-to-time and also depends strongly on the US system capa-bilities and settings.
The power and color flow Doppler images cansuffer from reverberations and other artificial motion
Fig. 4. Power Doppler image of blood vessel in finger with
flash artifact (left) and reverberation artifact (right).
artifacts, principally flash artifacts that must be elimi-nated for blood flow quantification as shown in Fig. 4.The flash artifacts arise mainly due to patient breathingmotion during the scan. The reverberation artifact couldbe exacerbated by our use of the thin TPX membrane inbetween the scanner and the human tissue for conve-nience and positioning purposes (refer to Fig. 1). Most ofthe artifacts seen in the power Doppler images had incommon a certain spatial pattern of reverberations thatrepeats itself along the axial direction.
A boundary was traced manually to outline the threefingers and all the color data outside this outline wereneglected. Taking a cross-section of the 3D gray-scaleimage at midaxial point did not give clear outer fingerboundary but the mean value of the spatial one-dimen-sional (1D) Fourier transform taken along the axial di-
Fig. 5. Schematic of projection of mean value of spatial 1DFFT of power Doppler onto the cross-plane (see Fig. 6).
Fig. 6. Mean value of the power of 1D spatial FFT of the power
Doppler signal helps in identifying the finger boundary.
396 Ultrasound in Medicine and Biology Volume 34, Number 3, 2008
rection in the Doppler window of power Doppler datagave clear boundary. In Fig. 5, a schematic outlining thecalculation of spatial 1D Fourier transform of the Dopp-ler data are shown. Fig. 6 has the mean-value of spatial1D Fourier transformed data in top-view with the fingeroutline marked for clarity.
The number of color pixels were summed withinthis finger boundary along the axial direction in eachimaging plane. As flash and reverberation artifacts areexpected to be of high amplitude and, due to their repet-itive nature along the axial direction, a threshold was setfor the color pixel count in terms of percentage of totalpixels along the axial direction. All color pixel valuesalong the axial line that exceeded the threshold wereneglected. This was evaluated for about eight trainingsets of Doppler data to determine the threshold value. Forpower Doppler data shown in Fig. 7(a), a threshold valueof 25% cuts out artifacts as well as some amount of colordata as in Fig.7(b), while 75% threshold allows almostall the artifacts pass through as shown in Fig. 7(d). Athreshold value of 50% that blocked most of the artifactswithout cutting the blood flow as shown in Fig. 7(c) was
Fig. 7. (a) Power Doppler image of finger that contains rever-beration artifact (note that the region outside the finger outlinewas blacked out), (b) after passing through an amplitude thresh-old filter of 25% that reduces both flash artifact and someamount of blood flow, (c) after threshold filter of 50% thatreduces flash artifact and (d) after threshold filter of 75% thatallowed most of the artifacts to pass through. Notice that 50%
threshold filter gives the best result.
used for Doppler analysis of all the data. This thresholdfilter works for both the flash and reverberation artifact.
The color pixel density is obtained by summing thepower Doppler signal in the post-threshold filtered 3Ddata and dividing by the volume of the selected region-of-interest (ROI) of finger. The color pixel density esti-mate acts as a measure of blood flow in the periphery andvery high changes in it could be used as an indicator ofhigh heart-rate or high stress. This could indicate that theindividual may not be acting in free-will.
RESULTS
Image registrationUpon registering pairs of image sets, a rotational
matrix (R) and scaling matrix (S) were obtained thatgives the amount of rotation and scaling along the threeaxes. The 3 � 3 R and S matrices were almost equal tothe identity matrix with a small deviation for data fromthe same person while they differ from the identitymatrix by a large margin for two different people. Wedefined the tolerance limit T, as the maximum allowedstandard deviation of diagonal elements from unity andthe nondiagonal elements from zero. For a successfulmatch, the T value had to be a reasonable value for boththe R and S matrices.
��i�1
3
�Rii � 1�2 � T &��i,j�1
3
�Rij�2 � T, for i � j
��i�1
3
�Sii � 1�2 � T &��i,j�1
3
�Sij�2 � T, for i � j
For T � 0.1, the image-based registration met criteria formatching in all 15 out of 15 scans on the same individualand failed to meet those criteria in all of the 12 sets offinger scans from different subjects giving a sensitivityand specificity of 1.00 and 1.00, respectively. For astricter tolerance level of T � 0.05 on the same data sets,we obtained 14 right matches out of 15 scans of the samesubject and 12 mismatches out of 12 scans from differentsubjects providing overall accuracy of 93% (sensitivity �0.93 and specificity � 1.00) .
Reader studyThe results of reader study of all 36 image pairs
from all four readers have been summarized in Fig. 8 inthe form of a scatter plot (a) and the correspondingreceiver operator characteristic curve (b). The area underthe receiver operating characteristic (ROC) curve is high(Az � 0.96) signifying high accuracy for detectingmatches and mismatches (Hanley et al. 1982, 1983).
Ultrasound of the fingers for human identification ● G. NARAYANASAMY et al. 397
ologists yielded better results with Az � 1.00 than that byother readers. This Az value is the same as for theprevious image-based registration result with the toler-ance limit T � 0.1.
Cardiac-gated Doppler studyAfter making the correction for reverberation arti-
facts in the color pixel density of the power Doppler dataon all subjects, the average fractional change in bloodvolume detected with exercise was �24% (Fig. 9) in allthe 31 power Doppler scans that were collected. Theaverage fractional change in blood volume detected was�43% from among the remaining 19 Doppler scans, ifthe five scans where change was small and nine scans
Fig. 8. (a) Scatter plot of all four reader grades for matched withunmatched cases. (b) Corresponding receiver operating curve(ROC) curve shows the probability of matches with the probability
of false alarm for two expert readers and all the four readers.
where change was negative were excluded (reasons pro-vided in Discussions and Conclusions). Fig. 9 shows thepercentage change in CPD with exercise.
DISCUSSIONS AND CONCLUSIONS
This finger study using 3D ultrasound illustrates thepotential of using internal structures and vascularity offingers and other biometrics for identification of individ-uals. The use of mutual information based image regis-tration software for human identification was illustratedon gray-scale US data. Future work should investigatethe use of an overlap invariant measure like normalizedmutual information (Studholme et al. 1999) and othermatching functions as an absolute metric of match. Thismight be obtained in conjunction with histogram equal-ization.
The human reader study accurately discriminatedindividuals from gray-scale US images. The Az value of1.00 with two musculoskeletal radiologists and 0.96 withall four readers suggests that with modest confirmingstudies, these techniques could be used to supplementconventional techniques, particularly as very inexpensiveUS finger imaging is becoming possible. Subjective as-sessment of the volar plate, tendon thickness and attach-ment site and bone contours enabled fingers from indi-viduals to be matched. While anechoic joint fluid in thevolar recess proximal to the volar plate may be present,this finding was not used for individual identification asjoint effusions may change over time. Some of the struc-tures investigated could change over time due to agingand osteoarthritis. Also, some finger structures in peoplehaving joint disorders like rheumatoid arthritis would bemore difficult to identify. The power Doppler imaging ofhuman finger flow brought out the variation of the blood
Fig. 9. Percent change in color pixel density (CPD) withexercise in 31 cases.
flow in an individual with time and stress. These poten-
398 Ultrasound in Medicine and Biology Volume 34, Number 3, 2008
tial variations must be controlled or accommodated inuse of the blood flow in the finger as a supplementalbiometric. Furthermore, such measurements were depen-dent on the US system capabilities and settings. Some ofthe challenges are addressed below.
Reasons for decrease in blood flow with exercise insome cases. Despite the short exercise routine, five out of31 power Doppler scans detected a very low volume ofblood flow. The blood flow could not be distinguished fromthe noise level and, hence, were neglected from this follow-ing analysis. The mean time required for repositioning fromthe end of exercise to the start of the Doppler US scan wasabout 1 min during which the heart-rate and blood pressurefell by various amounts. Care was required to minimize thisinterval to control this source of error.
An important reason for decrease in the blood flowcould be error in determining the pulse transit time(PTT), i.e., the time interval for the detectable blood flowto peak at the finger after the R-wave in the ECG signal.This made us drop seven cases where the color pixeldensity change was negative. With physical stress, anx-iety, body temperature, metabolism etc., there is a shift inan individual’s PTT to limbs (Furedy et al. 1996; Nas-chitz et al. 2003). As we performed cardiac-gated Dopp-ler measurements at a fixed PTT (250 ms) to reduce theexam time, considerable error in timing the externaltrigger to the actual peak pulse arrival time might haveresulted (Nitzan et al. 2002). A better approach nowavailable on some commercial ultrasound scanners is thecolor peak detect mode, which could be set to record thehighest Doppler signal power over a cardiac cycle.
A pilot study was conducted to measure the varia-tion of PTT on both left and right hands with changes in
Fig. 10. Variation of pulse transit time (PTT) at finger (in ms)with heart-rate (per min) of a volunteer on the index and middle
finger on the left and right hands.
heart-rate caused by physical exercise and Fig. 10 pre-sents the results. The exercise routine used here (joggingup the stairs for 4 min) was harder than the one used forthe finger study but it highlights the fact that PTTchanges considerably in an individual with physical ex-ercise. The mean PTT in the middle and index fingers inthe right and left hands of the volunteer changed from246 to 189 ms in this study.
In conclusion, 3D US scanning of the finger offersa novel biometric for human identification and basicphysiological status. This study illustrated identificationof individuals by matching of their finger scans taken atdifferent times when compared with scans of other indi-viduals. This matching was accomplished by an objec-tive, semiautomated algorithm based on the mutual in-formation metric and by a human reader visual study.More development is required to make such imaging areliable, high speed tool for detecting an individual fromamong millions. The technique is close to utility in someapplications with reasonable assurance that an individualis the one in question, such as in high value cell phonesecurity, particularly as an adjunct to fingerprint analysisthat can be misinterpreted easily. The color Dopplerstudy highlights issues with blood flow variation in anindividual that need to be addressed for using blood flowas a potential biometric marker.
Acknowledgments—This work was supported in part by NIH grantsPO1 CA87634 and R01 CA91713 and by NIST Cooperative Agree-ment No. 2001 to 00-4392 through CrossMatch Technologies Inc.Moreover, we thank Timothy D. Johnson (Asst. Professor, School ofPublic Health, UM) for his help and Dr. Zhi Yang (Research Fellow,Department of Radiology, UM) for serving as a reader.
REFERENCES
Bhatti PT, LeCarpentier GL, Roubidoux MA, Fowlkes JB, Helvie MA,Carson PL. Discrimination of sonographically detected breastmasses using frequency shift color Doppler imaging in combinationwith age and gray scale criteria. J Ultrasound Med 2001;20:343–350.
Carson PL, Li X, Pallister J. Approximate quantification of detectedfractional blood volume in the breast by 3D color flow and Dopplersignal amplitude imaging. Proc IEEE Int Ultrason Symp 1993; Cat.#93CH3301-9:1023–1026.
Entrekin R, Jackson P, Jago JR, Porter BA. Real time spatial compoundimaging in breast ultrasound: Technology and early clinical expe-rience. Medica Mundi 1999;43:35–43.
Furedy J, Szabo A, Peronnet F. Effects of psychological challengers onheart rate, T-wave amplitude and pulse-transit time. Int J Psycho-physiol 1996;22:173–183.
Giger ML, Al-Hallaq H, Huo Z, Moran C, Wolverton DE, Chan CW,Zhong W. Computerized analysis of lesions in US images of thebreast. Acad Radiol 1999;6:665–674.
Golub RM, Parsons RE, Sigel B, Feleppa EJ, Justin J, Zaren HA, RorkeM, Sokil-Melgar J, Kimitsuki H. Differentiation of breast tumors byultrasonic tissue characterization. J Ultrason Med 1993;12:601–608.
Hanley JA, McNeil BJ. The meaning and use of the area under a ReceiverOperating Characteristic curve. Radiology 1982;143:29–36.
Hanley JA, McNeil BJ. A method for comparing the areas under the
Receiver Operating Characteristic cases derived from the samecases. Radiology 1983;148:839–843.
Ultrasound of the fingers for human identification ● G. NARAYANASAMY et al. 399
Jain AK, Hong L, Pankanti S. An identity-authentication system usingfingerprints. Proc IEEE 1997;85:1365–1388.
Jain AK, Pankanti S, Prabhakar S, Hong L, Ross A, Wayman JL.Biometrics: A grand challenge. Proc Int Conf Pattern Recognition(ICPR) 2004;2:935–942.
Meyer CR, Boes JL, Kim B, Bland PH, Zasadny KR, Kison PV, KoralK, Frey KA, Wahl RL. Demonstration of accuracy and clinicalversatility of mutual information for automatic multimodality im-age fusion using affine and thin-plate spline warped geometricdeformations. Med Image Anal 1997;1:195–206.
Naschitz JE, Itzhak R. Assessment of cardiovascular reactivity byfractal and recurrence quantification analysis of heart-rate and pulsetransit time. J Hum Hypertens 2003;17:111–118.
Neemuchwala H, Hero AO, Zabuawala S, Carson PL. Image registra-tion methods in high-dimensional space. Intl J Imaging Sys Tech2006;16(5):130–145.
Nitzan M, Khanokh B, Slovik Y. The difference in pulse transit time tothe toe and finger measured by photo-plethysmography. Physiol
Meas 2002;23:85–93.
Pankanti S, Prabhakar S, Jain AK. On the individuality of fingerprints.Proc IEEE Comput Soc Conf Comput Vis Pattern Recognition2001;1:805–812.
Park H, Bland PH, Brock KK, Meyer CR. Adaptive registration usinglocal information measures. Med Image Anal 2004;8:465–473.
Rasband WS, Image J, US National Institutes of Health, Bethesda, MD,USA. Available at: http://rsb.info.nih.gov/ij/, 1997–2007. AccessedSeptember 1, 2007.
Rubin JM, Adler RS, Fowlkes JB, Spratt S, Pallister JE, Chen JF,Carson PL. Fractional moving blood volume estimation usingDoppler power imaging. Radiology 1995;197:183–190.
Sahiner B, Chan HP, Roubidoux MA, Helvie MA, Hadjiiski LM,Ramachandran A, Paramagul C, LeCarpentier GL, Nees A, BlaneC. Computerized characterization of breast masses on three-dimen-sional ultrasound volumes. Med Phys 2004;31:744–754.
Studholme C, Hill DJG, Hawkes DJ. An overlap invariant entropymeasure of 3D medical image alignment. Pattern Recognition
