Predictive medicine: Computational techniques in...

Biomedical Paper

Predictive Medicine: Computational Techniquesin Therapeutic Decision-Making

Charles A. Taylor, Ph.D., Mary T. Draney, M.S., Joy P. Ku, M.S., David Parker, B.S.,Brooke N. Steele, M.S., Ken Wang, M.S., and Christopher K. Zarins, M.D.

Division of Vascular Surgery, Department of Surgery (C.A.T., C.K.Z.), Department of MechanicalEngineering (C.A.T., M.T.D., D.P., B.N.S.), and Department of Electrical Engineering

( J.P.K., K.W.), Stanford University, Stanford, California, USA

ABSTRACT The current paradigm for surgery planning for the treatment of cardiovasculardisease relies exclusively on diagnostic imaging data to define the present state of the patient,empirical data to evaluate the efficacy of prior treatments for similar patients, and the judgement ofthe surgeon to decide on a preferred treatment. The individual variability and inherent complexityof human biological systems is such that diagnostic imaging and empirical data alone are insufficientto predict the outcome of a given treatment for an individual patient. We propose a new paradigmof predictive medicine in which the physician utilizes computational tools to construct and evaluate acombined anatomic/physiologic model to predict the outcome of alternative treatment plans for anindividual patient. The predictive medicine paradigm is implemented in a software system developedfor Simulation-Based Medical Planning. This system provides an integrated set of tools to testhypotheses regarding the effect of alternate treatment plans on blood flow in the cardiovascularsystem of an individual patient. It combines an internet-based user interface developed using Java andVRML, image segmentation, geometric solid modeling, automatic finite element mesh generation,computational fluid dynamics, and scientific visualization techniques. This system is applied to theevaluation of alternate, patient-specific treatments for a case of lower extremity occlusive cardiovas-cular disease. Comp Aid Surg 4:231–247 (1999). ©1999 Wiley-Liss, Inc.

INTRODUCTIONSignificant advances have been made in the diag-nosis and treatment of cardiovascular disease sincethe introduction of modern medical imaging tech-nology and surgical and pharmacologic therapeuticoptions. In recent years, three-dimensional (3D)cardiovascular imaging techniques and medical vi-sualization software have enabled physicians tointeractively view the cardiovascular anatomy ex-ternally and even to “fly through” blood vessels inorder to examine sites of disease.27,30,37–40In addi-tion to obtaining anatomic data, physicians can

obtain physiologic data from a variety of sources,including Doppler ultrasound and magnetic reso-nance imaging. It is now possible to obtain time-resolved 3D flow fields in the cardiovascular sys-tem using magnetic resonance imaging (MRI)techniques in a matter of minutes.14 In parallel withthese developments in diagnostic methods, thera-peutic options including conventional surgical re-pair, minimally invasive endoluminal approaches,and drug and gene therapy have emerged. Withinthese different classes of therapeutic options liemany more subclasses. The physician must choose

Received March 5, 1999; accepted August 20, 1999.

Address correspondence/reprint requests to: Charles A. Taylor, Ph.D., Division of Vascular Surgery, Department of Surgery,Stanford University, Stanford, CA 94305, U.S.A. Telephone: (650) 723-3487; Fax: (650) 723-3521. E-mail:[email protected].

Computer Aided Surgery 4:231–247 (1999)

©1999 Wiley-Liss, Inc.

between a variety of devices and pharmacologicagents, in addition to choices amongst differentclasses of therapeutic options.

The multitude of therapeutic choices whichare available to a physician has led to a complexdecision-making process. At present, this decision-making process is based largely on experience andempirical data, and involves predicting the outcomeof a selected therapy for an individual patient usinginformation from “similar” therapies applied previ-ously to other “similar” patients. This process canbe enhanced with statistical decision support tools,but these tools still rely primarily on empirical datawhich might not reflect the condition of that indi-vidual subject.16 In essence, this decision-makingprocess is rich with information about the presentand the past for other subjects, but lacks informa-tion or predictions about the future for the individ-ual patient.

In a wide range of disciplines, predictions offuture states are made on the basis of simulationtechniques and theoretical models initialized withdata from the past and present. Increasingly, thesesimulation methods are implemented on computers.In the engineering disciplines the use of computersis often divided into three distinct areas: Computer-Aided Design, Computer-Aided Engineering, andComputer-Aided Manufacturing. The role of sim-ulation is most prominent in Computer-Aided En-gineering, where mathematical models are imple-mented using numerical methods to evaluatedesigns developed using Computer-Aided Designsoftware. After an appropriate design is identifiedusing conventional and Computer-Aided Engineer-ing methods, Computer-Aided Manufacturingmethods are used to control the machinery used tofabricate a product.

In medicine, two significant recent develop-ments in the application of computer technologyare computer-aided surgical planning and image-guided surgery. Contemporary computer-aided sur-gical planning software is analogous to Computer-Aided Design tools and image-guided surgerydevelopments are analogous to Computer-AidedManufacturing tools. These tools are largely anat-omy-based, e.g., creating anatomic models fromimaging data to address issues of surgical approachand to aid the surgeon in implementing a treatmentplan. In the field of neurosurgery these techniquesare aiding surgeons in repairing cerebral aneurysmsand arterio-venous malformations, and in resectingtumors.17,21,29,42 Similar methods are being em-ployed for a wide range of applications includingcraniofacial, orthopedic, lung, prostate, and liver

surgery planning.13,52 However, surgery planningmust not only address questions of surgical ap-proach but also expected outcome. The predictionof future-states, e.g., the efficacy of a treatmentoption for restoring blood flow or the performanceof a device implanted in the cardiovascular system,requires a procedure for simulating physiologicfunction in addition to the diagnostic imaging datadefining the anatomy of the individual patient. Thefundamental shortcoming of contemporary com-puter-aided surgical planning methodology is that itlacks thesimulation tools analogous to the Com-puter-Aided Engineering tools used in engineeringapplications.

For the cardiovascular system, the simulationof physiologic function includes the modeling ofthe complex fluid mechanical behavior of blood.Fortunately, significant progress has been made inunderstanding and modeling cardiovascular bloodflow. Hemodynamic, or blood fluid mechanical,conditions, including velocity, shear stress, andpressure, play an important role in the modulationof vascular adaptation and the localization of vas-cular disease.15,23,53–55In recent years, computersimulation software has enabled scientists to con-struct and evaluate models of human anatomy andphysiology and test hypotheses regarding the roleof hemodynamic factors in vascular adaptation anddisease.25,34–36,45–48As these investigations are de-signed to be compared with patterns of adaptationand disease for the population as a whole, theseinvestigations are typically conducted based on ide-alized computer models representing “average”anatomy and physiologic conditions.

In addition to the investigation of the role ofbiomechanical factors in adaptation and disease,computer simulation technology can also be ap-plied to the design of surgical procedures, e.g., theconstruction of anastomoses with improved hemo-dynamic conditions24 or the evaluation and designof orthopedic procedures.4–11 These investigationsare typically performed based on idealized, notpatient-specific, models, but can provide substan-tial quantitative information and improve surgicalpractice.

In the predictive medicine paradigm proposedherein, a physician would use diagnostic data toreconstruct a model of an individual’s anatomy andphysiology, and then use simulation techniques,implemented in a simulation-based medical plan-ning software system, to predict the response ofthat patient to alternate treatments under differentphysiologic states. Clearly, the success of predic-tive medicine techniques is dependent upon how

232 Taylor et al.: Predictive Medicine

faithfully the mathematical model reflects the ac-tual anatomy and physiology of the individual, andon how efficiently the simulations can be per-formed. For blood flow in the human vascularsystem, the patient-specific anatomic model is de-rived from 3D imaging techniques including CTand MRI. The patient-specific physiologic modelis, in comparison to the anatomic model, an ab-stract model based on the equations governingblood flow in arteries and on patient-specific diag-nostic data, e.g., measurements of blood flow andpressure. Once the patient-specific model is cre-ated, the predictions of the consequences of treat-ments can be made based on the simulation ofblood flow in models that reflect the proposed treat-ments. This simulation of blood flow can be per-formed using techniques from computational fluiddynamics on high-performance computers.46

A software system developed for simulation-based medical planning is introduced. This systemprovides an integrated set of tools to test hypothe-ses regarding the effect of alternate treatment planson vascular hemodynamics, and was applied to aclinical case of lower extremity vascular disease topredict the effect of multiple alternate treatmentplans on blood flow and pressure.

MATERIALS AND METHODSWe have created a comprehensive software frame-work for simulation-based medical planning. Thissoftware system, ASPIRE (Advanced SurgicalPlanning Interactive Research Environment), hasan internet-based user interface developed usingthe Java programming language and the VirtualReality Modeling Language (VRML). In addition,this system includes image segmentation, geomet-ric solid modeling, automatic finite element meshgeneration, computational fluid dynamics, and sci-entific visualization techniques.

Custom visualization software was developedfor the visualization and segmentation of medicalimaging data.31,32 A geometric modeler was em-bedded in this visualization environment to allowthe user to interactively fit curves to sets of pointsrepresenting the cross-section of a blood vessel,group and skin these curves to create a NURBsurface, then cap and bound this surface to create asolid object of a single blood vessel.51 The individ-ual blood vessels are then unioned together to forma single solid model of a portion of the vascularsystem. This solid model, generated from patient-specific imaging data, represents the patient’s vas-cular anatomy prior to treatment. This preoperativemodel can then be modified by enlarging vessels to

simulate angioplasty and adding vascular bypassgrafts using solid modeling Boolean procedures.The pre- and post-operative model is then dis-cretized using an automatic finite element meshgenerator.43

In general, the model of the patient’s physi-ology includes the governing equations for bloodflow, boundary and initial conditions, and constitu-tive properties of the blood and vessel wall. As afirst approximation, the blood is treated as a New-tonian fluid and the blood vessels are treated asbeing rigid. The boundary conditions for the sim-ulations can be obtained using patient-specific di-agnostic data, e.g., MRI measurements of bloodflow.

Once the patient-specific model is created, aprediction of the consequences of treatments re-quires the simulation of blood flow in models thatreflect the proposed treatments. The finite elementmethod is employed in the present investigation, asit is well suited to the complex geometric modelsencountered in vascular modeling. The finite ele-ment method employed is based on the theory ofstabilized finite element methods developed byHughes and colleagues.2,18,19 The method is de-scribed in detail by Taylor et al.46 as it is imple-mented in the commercial finite element programSpectrum.3 The computational method employedhas been shown via analytical and experimentalvalidation studies to yield accurate solutions forpulsatile flow problems.46

Another critical issue in the application ofcomputational techniques to surgical planning isthe visualization and interpretation of the computedresults. For blood flow, alternative techniques tovisualize flow fields include velocity vectors,streamlines, particles, and isosurfacing techniques,as well as contouring techniques for scalar fieldssuch as pressure. In addition to scientific visualiza-tion techniques, simple x-y plots are essential forinterpreting computed results. Figure 1 describesthe overall procedure for simulation-based medicalplanning, starting with the acquisition of patientdata and ending in a comparison of alternate treat-ments.

A typical surgery planning session starts byexamining a case presentation. Patient history,medical imaging data, and vascular lab data areexamined through a web browser, as shown inFigure 2 for a case of lower extremity occlusivedisease. A pre-operative 3D geometric model of theblood vessels can also be seen in Figure 2.

The next step in the surgery planning processis the specification of a treatment plan. This is done

Taylor et al.: Predictive Medicine 233

by drawing a treatment on a surgical sketchpad, asshown in Figure 3. This interactive functionality isimplemented using the Java programming languageand allows the user to specify the starting andending locations of vascular bypass grafts. Oncethe graft is drawn on the sketchpad, a menu appearsto allow the user to specify graft characteristics,including dimensions, material, and manufacturer.At present, this data is used in an external programto create a new “postoperative” geometric solidmodel, e.g., by unioning the preoperative anatomicmodel with a vascular bypass graft.46

Once the treatment plan is constructed, thenext step is to perform calculations to simulate theblood flow in the postoperative model. The geo-metric model is discretized using automatic finiteelement mesh generation techniques.43 Physiologicboundary conditions are established based on amodel of flow resistance derived from pre-opera-tive data and calculations. Blood flow in the post-operative models is computed under a variety ofphysiologic states, including resting and exerciseconditions.47,48 Once the blood flow analyses havebeen performed, the results are viewed by using thesystem in the treatment evaluation mode. As shownin Figure 4, quantitative results for blood pressure,flow velocity, and volume flow rate can be exam-ined in a set of graphical display windows. Thesegraphical tools can be used to compare the blood

flow before and after a treatment plan or to com-pare the results of two alternate plans. In addition tothe graphical display windows, the surgeon canvisualize 3D models of the preoperative or postop-erative geometry, pressure, or blood flow using thetwo VRML windows.

CLINICAL CASEIn order to test the concept of predictive medicine,we created a “mock” clinical case and applied oursimulation-based medical planning methodology.The base anatomy for this case was generated byextracting the luminal boundaries from a normalsubject. This subject was injected with a contrastagent (Gd-DPTA) and then imaged using a phase-contrast MR angiographic sequence from the chestdown to the feet with a 1.5 T GE Signa system.1 Atotal of 384 slices were acquired in three scans of128 slices, with in-plane resolution of 0.9375 mmand slice thickness of 1.8 mm.

The mock clinical case created from this baseanatomy is that of a 57-year-old man who presentswith claudication pain in his right hip and buttocks.Upon physical examination, a blue big toe andnon-healing ulcerations are noted on the right foot.The prior medical history includes unsuccessfulangioplasties of the right iliac artery and right su-perficial femoral artery. A schematic summarizingthe result of diagnostic imaging is shown in Figure5, where the right iliac artery is noted to be com-pletely occluded, the left iliac artery has a 50%reduction in diameter, the right superficial femoralartery has a diffuse 50% reduction in diameter, andthe left superficial femoral artery is completelyoccluded. There are many different surgical andnon-surgical treatments which could be attemptedto correct this patient’s clinical problems, but anysuccessful procedure would have to increase theblood flow to the right leg, in order to eliminate thepain experienced upon walking, and to the rightfoot if the wound is to heal. The pre-operativegeometric model and three of the most commonprocedures, given this particular patient’s clinicalpresentation and observed disease, are shown inFigure 6. These procedures were modeled and com-pared to see which treatment would produce thebest hemodynamic outcome for this specific case.A finite element mesh was generated for each of thegeometric models. Figure 7 shows the surface ofthe finite element mesh generated for the case ofthe aorto-femoral bypass with a proximal end-to-side anastomosis of the graft to the aorta (Fig-ure 6b).

Resting flow conditions were used to assess

Fig. 1. Flowchart of overall procedure for Simulation-Based Medical Planning.


the blood flow in the foot needed for wound heal-ing. Exercise flow conditions were used to assessthe blood flow in the right leg under walking con-ditions. The boundary conditions for pre-operativeand post-operative computations were prescribedas follows. First, pre-operative analyses under rest-ing and exercise steady flow conditions were per-formed with a specified volume flow rate througheach boundary based on literature data for flowdistribution.20 This pre-operative analysis was usedto compute the average pressure distribution ateach outflow boundary. Second, a unique resistancevalue was computed for each outflow boundarybased on a relationship between pressure and vol-ume flow rate of the formP 5 QR, whereP is themean pressure,Q is the volume flow rate, andR isthe resistance to flow. For each outflow boundary,

the same resistance value was used for all of thepost-operative analyses for each surgical plan. Us-ing this strategy, the volume flow rateQ and pres-sureP were calculated (not specified) for each ofthe boundaries for each of the surgical plans. Thenonlinear evolution equations governing bloodflow for each treatment plan were then solved forthe velocity and pressure fields over 3 cardiac cy-cles with 200–300 time steps per cardiac cycleusing 8 processors of a Silicon Graphics Origin2000® parallel computer.

This system for Simulation-Based MedicalPlanning was demonstrated at the 1998 Society forVascular Surgery (SVS) meeting. For this demon-stration, this system was applied to predict theeffect of multiple alternate treatment plans onblood flow and pressure under resting and exercise

Fig. 2. ASPIRE user interface in clinical data mode running in the web browser Netscape. Angiographic data is shown inthe main window for a patient with lower extremity occlusive disease. Shown in the two panels on the right are 3D modelsconstructed from pre-operative MRI data. Pre-operative anatomy and physiology data can be examined and visualized usingthe Virtual Reality Modeling Language (VRML).


conditions for the clinical case of lower extremityvascular disease described above. Four prominentvascular surgeons (all former presidents of theSVS) were selected based on their clinical experi-ence; three of the four surgeons had little priorcomputer experience. All of the surgeons weregiven 15–20 minutes of training the day prior to thedemonstration. At the start of the demonstration,the surgeons were presented with the clinical caseand asked to use the system to implement andevaluate one of four possible surgical plans. Thedisplay from each of the SGI Octane® computersused by the vascular surgeons was projected to alarge display screen for viewing by the audience ofvascular surgeons and affiliated health care profes-sionals.

RESULTSFigure 8 shows the pressure contours at peak sys-tole for the preoperative model and the three treat-ment plans under resting conditions. An ideal pro-cedure would have minimal pressure losses totransport blood from the aorta to the peripheralvessels. The aorto-femoral bypass procedure with a

proximal end-to-side anastomosis (Fig. 8b) resultedin the lowest pressure losses of any procedure. Theaorto-femoral bypass procedure with a proximalend-to-end anastomosis (Fig. 8c) had similarly fa-vorable results at the level of the popliteal arteries,but had a significant pressure loss for blood flowingretrograde through the left iliac artery stenosis tothe distal aorta. The angioplasty with femoral-to-femoral bypass graft (Fig. 8d) had the greatestpressure losses of any procedure. Also evident forthis procedure is the significant pressure loss fromthe aorta to the left femoral artery and then acrossthe femoral-femoral bypass graft.

Figure 9 depicts the pressure waveformsunder resting conditions at right and left femoralarteries for the pre-operative and post-operativecases. Note the relatively low pressures in theright femoral artery pre-operatively, and the factthat the angioplasty with femoral-to-femoral by-pass graft has lower pressures in the right fem-oral artery than either aorto-femoral bypass pro-cedure. In the left femoral artery, the pressuresare elevated for the aorto-femoral bypass proce-dures and compared to the pre-operative pressure

Fig. 3. ASPIRE user interface in treatment planning mode. Shown on the far left are a variety of buttons to query the modeland implement treatment options. These treatment options are sketched on a 2D projection of a pre-operative model. The redportion of the model represents the luminal surface of the blood vessels. The yellow portion represents the location where anangioplasty is to be performed to dilate a diseased blood vessel. The light blue line represents the locations where afemoral-to-femoral bypass is to be introduced. A graphical panel is shown to the right of the surgical plan. This panel allowsfor the extraction of quantitative comparisons of pressure, velocity, and volume flow at standard or user-defined measurementlocations.


waveform. For the angioplasty with femoral-to-femoral bypass graft, the peak systolic pressureis predicted to actually decrease compared to thepre-operative peak systolic pressure due to thesubstantially increased left femoral artery flowresulting from this procedure.

Table 1 details a comparison between thecomputed mean flow rates under resting conditionsfor the pre-operative simulation and the three treat-ment plans. Note, in particular, the mean flow ratesfor the alternate treatment plans at the levels of theaortic bifurcation and the right popliteal artery. Forthe case of the aorto-femoral bypass procedure withthe end-to-end anastomosis, the flow at the level ofthe aortic bifurcation is exactly the flow that willperfuse the inferior mesenteric and lumbar arteries.

It is noted that the flow rate at the level of the aorticbifurcation is only 15 ml/min for this case. Inaddition, in all cases, the flow increases in the rightpopliteal artery compared to the pre-treatment case,with the greatest improvements in flow seen for theaorto-femoral bypass procedures.

The demonstration at the 1998 Society forVascular Surgery (SVS) meeting was completedwith no technical difficulties encountered. Each ofthe four vascular surgeons on the panel success-fully used the software system to solve the clinicalcase presented in this live demonstration. Figure 10shows the panelists using the software system aswell as the display from one of the computersprojected on a large screen for viewing by theaudience.

Fig. 4. ASPIRE user interface in treatment evaluation mode. The increase in flow at the level of the right popliteal arteryis examined quantitatively by comparing the pre-operative volume flow to the post-operative volume flow. The panel at theright allows the qualitative comparison of the pressure contours under resting conditions for the pre-operative case and theimplemented surgical procedure.


DISCUSSIONA software system developed for Simulation-BasedMedical Planning was introduced and applied toexamine the effect of different treatment plans onblood flow and pressure for a case of lower extrem-ity vascular disease. The significance of the clinicalfindings is discussed, followed by an analysis of themodeling assumptions and techniques employed.Finally, the role of simulation techniques in medi-cine of the future is considered.

The computational analyses of blood flow forthe case of lower extremity occlusive disease gen-erated a wealth of information regarding blood flowand pressure for the treatments examined. It mustbe emphasized that these findings are patient-spe-cific, and could be different for a different subjectwith different anatomy or physiologic conditions.Clearly, clinical studies will be necessary to test thevalidity of these findings, and, as in the case of new

theoretical tools introduced in other fields, new ex-perimental studies will be initiated to test predictions.

Of particular interest to the vascular surgeonson the SVS panel was the prediction that the aorta-femoral bypass plan with a proximal end-to-endanastomosis (AFB w/ E-E) had a very low flow rate(15 ml/min) at the level of the aortic bifurcationunder resting conditions (see Table 1 and also notethe lower pressures at the level of the aortic bifur-cation in Figure 7c). This is of concern since forthis particular case it would be expected to lead toa significantly reduced blood flow to the lowerintestine and colon. This finding might lead thevascular surgeon to reimplant the inferior mesen-teric artery into the side of the bypass graft. Alter-natively, some vascular surgeons have suggestedthat this finding might lead them to perform anangioplasty on the left common iliac artery whenperforming this procedure to ensure adequate ret-rograde flow to the inferior mesenteric artery.

An additional finding which generated con-siderable discussion at the SVS meeting was theresult that the treatment plan with the angioplastyand femoral-to-femoral bypass graft had only 50%of the flow at the level of the right popliteal arterycompared to the two procedures with aorto-femoralbypass grafting (see Table 1). Ultimately, the an-gioplasty with femoral-to-femoral bypass graft stillrequires blood flow for two legs to flow down oneiliac artery. With quantitative, patient-specific in-formation at their disposal, the vascular surgeoncould weigh the increased trauma, yet improvedblood flow, of the aorto-femoral bypass procedureswith the less invasive nature, yet inferior bloodflow, of the angioplasty and femoral-to-femoralbypass procedure. Once again, the conclusiondrawn is patient-specific, and the key point is thatthe surgeons will have quantitative predictions ofoutcomes to augment their professional experiencein the decision-making process.

Simulation-based medical planning requiresthe creation of patient-specific anatomic and phys-iologic models. The complexity of the cardiovas-cular system requires that assumptions be made forthese models. In regards to the individual patient’svascular anatomy, it is assumed that only the majorarteries which can be detected with diagnostic im-aging methods need to be included in the compu-tational model. This is justified based on the factthat the physician intervenes to restore blood flowonly in the major arteries and does not treat thesmall vessels. The effect of the small vessels in thedownstream vasculature is modeled through theboundary conditions. The patient-specific physio-

Fig. 5. Schematic illustrating result of diagnostic imagingfor patient with nonhealing wound in right foot.


logic model and pre-operative flow distribution wasdefined based on literature data for flow distribu-tions.20 In general, the pre-operative flow distribu-

tion will be defined based on pre-operative imagingdata, e.g., volume flow rates obtained using phase-contrast MRI techniques.

Fig. 6. Geometric models for alternate treatment plans for the case of lower extremity occlusive disease examined: (a)anatomic description with stenoses and occlusions shown, (b) aorto-femoral bypass graft with proximal end-to-side anasto-mosis, (c) aorto-femoral bypass graft with proximal end to end anastomosis, (d) balloon angioplasty in left common iliac arterywith femoral-to-femoral bypass graft.


The key software elements of the ASPIREsimulation-based medical planning system are theinternet-based user interface, the image segmenta-tion methods, geometric solid modeling, automaticfinite element mesh generation, boundary conditionspecification, computational fluid dynamics, andscientific visualization techniques. The authorshave developed and implemented methods for eachof these component technologies, but further im-provements are needed before the system can beused clinically. These component technologies, andthe assumptions which they employ, are discussedin turn.

The user interface to the ASPIRE system usesJava and the Virtual Reality Modeling Language(VRML) to ensure that the system works in adistributed computing environment. This enablesinteraction with models and simulation resultsacross a network from a variety of platforms. Someof the limitations of the current user interface in-clude the fact that the sketch-pad operates in atwo-dimensional (2D) mode and must be supple-mented by auxiliary 3D data supplied by the user,since all treatment plans are inherently three-di-mensional. As noted previously, in the present sys-tem, the segmentation, geometric modeling, andcomputations are performed external to the userinterface. At present, it is not feasible to distributeall of the image segmentation and geometric modelconstruction steps of the treatment planning pro-cess. In particular, the creation of the pre-operativegeometric model, which involves interacting withvery large 3D image data sets, is currently notfeasible in a fully distributed mode. However, wehave been successful in implementing distributedgeometric modeling and finite element mesh gen-

eration techniques using the Java Native Interface(JNI) and the Remote Method Invocation (RMI)facility. It is probable that, after a pre-treatmentmodel is constructed and loaded into a database,interventions can be implemented using a distrib-uted, client-server paradigm. In addition, the visu-alization of post-operative results can be performedusing the present internet-based methods. Futureenhancements in the distributed capabilities of thissystem will be enabled by the developments under-way in the Internet2 program.

The 3D visualization of medical imaging datasets is performed using either volume rendering orsurface rendering techniques. Volume renderingmethods have significant application in allowingfor the visualization of anatomic features withoutthe segmentation of image data. However, in orderto extract quantitative data and construct mathe-matically accurate data models, surface modelingand rendering techniques are most appropriate. Forsurface rendering techniques, surfaces within thevolume are commonly approximated by a polygo-nal mesh (a set of contiguous 2D primitives), ex-tracted using threshold-based Marching Cubestechniques with connectivity criteria.26 Althoughsuitable for visualization, polygonal meshes do notprovide representations most appropriate for geo-metric quantification and analysis, and are often toodense for distributed visualization applications ofvascular geometry. An alternative approach to con-structing surface models is using B-spline surfacesskinned from cross-sectional data or B-splinepatches fitted directly to surface triangulations.22

These inherently smooth surfaces can be connectedtogether using geometric modeling Boolean opera-tors to obtain anatomic models of the vasculature.32

In addition, these surfaces can be defined withhighly compact data structures appropriate for dis-tributed visualization applications. In the presentcase, we have created vascular geometric modelsby lofting B-spline surfaces through circular andelliptical cross-sections fitted to contour data on 2Dslices.32,46 Recently, we have extended this 2Dsegmentation technique to include piecewise linearsurfaces lofted through piecewise linear contoursextracted using a level-set method.49 This new ap-proach has the advantage of more accurately rep-resenting the vessel boundaries, but has not yetbeen fully implemented within our software frame-work.

The finite element mesh generation tech-niques employed resulted in largely isotropic dis-cretizations, although curvature-based refinementtechniques were used to achieve element sizes

Fig. 7. Close-up of finite element mesh for aorto-femoralbypass with end-to-side anastomosis.


which were appropriate for smaller branch vessels.New finite element mesh generation techniques areneeded which are more suitable for the discretiza-tion of arterial geometric models. These new meshgenerators, incorporating curvature-based refine-ment and boundary layer mesh generation tech-niques, could result in finite element meshes whichrequire 5–10 times fewer finite elements for a givenblood flow solution accuracy, as compared to con-ventional methods based on isotropic mesh refine-ment. This will enable up to a 100-fold decrease insolution time for blood flow computations. Clearly,these new mesh generation methods are essential

for the clinical application of simulation-basedmedical planning systems.

One of the greatest difficulties in applyingpredictive methods for vascular surgery planning isthe specification of appropriate “post-operative”boundary conditions for the computational analy-ses. This problem arises from the fact that mostcomputational methods for modeling blood flowrequire the specification of flow rate or pressure asboundary conditions, while the purpose of usingcomputational methods for treatment planning isoften to predict changes in these quantities. Quitesimply, the effect of a cardiovascular intervention

Fig. 8. Pressure distribution under resting conditions at peak systole for (a) pre-operative model, (b) aorto-femoral bypassgraft with proximal end-to-side anastomosis, (c) aorto-femoral bypass graft with proximal end-to-end anastomosis, (d) balloonangioplasty in left common iliac artery with femoral-to-femoral bypass graft.


on flow rate or pressure is unknown prior to theprocedure being performed, and flow rate and pres-sure are thus inappropriate boundary conditions forpredictive modeling.

The approach we have adopted is to dividethe cardiovascular system into two parts, one rep-resenting the major arteries where more detailedflow information is desired or which will be altereddue to the cardiovascular treatment, and the otherrepresenting the vasculature upstream or down-stream of the first part. The second part is assumedto be outside of the region of interest and althoughit might be affected by changes in blood flow orpressure, it will not be directly altered by the car-diovascular treatment. The first part of the cardio-vascular system is then modeled using a 3D pa-tient-specific anatomic model and governing

equations formulated to represent 3D flow fields,vessel wall mechanics, and interactions with de-vices or pharmaceuticals. The second part of thecardiovascular system, that which is outside theprimary region of interest (the first part), is mod-eled using a simplified description of cardiovascu-lar function. In the present case, this simplifieddescription involves specifying a constant periph-eral resistance, R, for each outflow boundary. Fi-nally, the 3D model (the first part) is coupled to thesimplified model (the second part) by enforcing thecontinuity of pressure and volume flow rate at theinterface between these two domains. For thepresent case, this reduces to a linear relation of theform P 5 QR, whereP is the mean pressure,Q isthe volume flow rate, andR is the resistance toflow. The simple constant resistance approach im-

Fig. 9. Pressure (mm Hg) vs. time (sec) over 1 cardiac cycle under resting conditions at (a) right and (b) left femoral arteriesfor pre-operative model (preop), balloon angioplasty in left common iliac artery with femoral-to-femoral bypass graft(femfem), aorto-iliac bypass graft with proximal end-to-end anastomosis (e2e), aorto-iliac bypass graft with proximalend-to-side anastomosis (e2s).


plemented thus far is entirely adequate for thesteady flow conditions, but has limited applicabilityfor pulsatile flow where the resistance will be afunction of time. Current efforts are focused on theimplementation of vascular network models basedon one-dimensional wave propagation theory. Ad-ditionally, the changes in the distal resistance dueto autoregulatory factors will need to be included infuture versions of this vascular surgery planningsystem.

The effect of vessel compliance was not con-sidered in the present investigation. Prior investi-gations of flow in deformable models have shownthat, in general, incorporating the effect of compli-ance does not significantly change the flowfields.12,34,44 It should be noted that, although ne-glecting wall distensibility may not have a majoreffect on the primary flow field, the incorporationof wall mechanics in these studies is important forother reasons, including the description of the stress

environment within the vessel walls and the inter-action between deformable vessels and vasculardevices. In summary, the use of rigid models forsimulation-based medical planning should beviewed only as a first approximation.

A Newtonian constitutive model for viscositywas employed in the present investigation. It isgenerally accepted that this is a reasonable firstapproximation to the behavior of blood flow inlarge arteries. Perktold et al.33 examined non-New-tonian viscosity models for simulating pulsatileflow in carotid artery bifurcation models. Theyconcluded that the shear stress magnitudes pre-dicted using non-Newtonian viscosity models re-sulted in differences on the order of 10% as com-pared with Newtonian models. Once again, the useof Newtonian models for blood rheology should beviewed only as a first approximation. Future stud-ies should incorporate non-Newtonian rheologicmodels.

Fig. 10. Photograph of demonstration illustrating panel of surgeons on-stage and projection to audience.

Table 1. Mean Flow Rate (ml/min)

Artery Pretreatment AFB w/E-S AFB w/E-EAngioplastyw/Fem-Fem

Abdominal Aorta 3009 3460 3460 3460Aortic Bifurcation 670 307 15 1062Right Internal Iliac 134 156 212 144Left Internal Iliac 464 148 150 335Right Femoral 216 615 663 444Left Femoral 464 550 781 747Right Popliteal 30 283 309 162Left Popliteal 272 351 362 290


Computational performance is a significantissue in deploying simulation-based medical plan-ning systems. With current software algorithms andhardware capabilities, the present calculations re-quired approximately 20 hours of CPU time peranalysis using 6 subdomains and 8 processors on aSilicon Graphics Origin 2000 parallel computer.The present computations require the simultaneoussolution of approximately 500,000 nonlinear equa-tions which must be solved at hundreds of timepoints per cardiac cycle. It is anticipated that, withincreased software and hardware performance,more complex problems will be solved. The com-peting demands of solution accuracy and computa-tional efficiency will have to be reconciled with thetime and economic constraints of the surgery plan-ning process. For many interventions, days orweeks can elapse between the times when imagingdata is acquired and when a treatment is imple-mented, but clearly it would not be economicallyviable to expend weeks of computational time intreatment planning for a given case. However, it islikely that 1–2 hours of total time expended for agiven patient and the evaluation of multiple treat-ment plans would be economically feasible formany cases. This will be achievable in the nearfuture with improvements in computer software,hardware, and networking.

Computational investigations of blood flowrequire validation using analytical solutions andinvitro and in vivo experimental methods. We havevalidated our finite element methods to modelblood flow in arteries using analytical solutions forpulsatile flow,50 and with laser Doppler anemom-etry velocity data obtained by Loth28 for steady andpulsatile flow in anin vitro model of an end-to-sideanastomosis.46 It was observed in these studies thatthe computational results compared favorably qual-itatively and quantitatively with the analytical so-lution and experimental data (less than 10% error atmost points). The present computations were per-formed with the same degree of finite element meshrefinement as these prior validation studies. Vali-dation studies in progress includein vitro studiesin deformable vessel models andin vivo studiescomparing subject-specific computational solu-tions to time-resolved 3D phase-contrast MRIvelocity data.

Clinical validation studies are essential forthe development and ultimate utilization of Simu-lation-Based Medical Planning systems. This isparticularly true considering the number of as-sumptions that are required to model blood flow inthe human cardiovascular system. Some of the crit-

ical points to consider in regard to the patient-specific anatomic models include an assessment ofthe accuracy in reconstructing anatomic modelsfrom imaging data, the accuracy with which staticmodels represent the actual, dynamic, human anat-omy, and the extent of the vascular anatomy thatneeds to be included in these models. The creationof the abstract patient-specific physiology modelsis more difficult than the creation of patient-specificanatomic models in that it is easier to conceptualizehow a treatment will alter the anatomy, but notobvious how a treatment will alter the physiologicresponse. As noted previously, this relates to theboundary conditions imposed on the computationalflow solutions. Extensive experimentation will berequired to identify and validate patient-specificphysiologic models. Methods for estimating theerror in the predictive simulations will be requiredas the uncertainty in the input conditions will bereflected in the uncertainty in the results. Stochasticmethods will ultimately be essential ingredients inany Simulation-Based Medical Planning system.

It is acknowledged that contemporary meth-ods for simulating blood flow in the cardiovascularsystem are crude compared to the true complexityand elegance of human function. However, thesetools provide a means for evaluating treatmentplans for individual patients which are simply notpossible without predictive methods. We believethat these predictive methods will add value to thecontemporary treatment planning process, but ulti-mately this question will have to be answered byclinical trials.

It is important to note that the informationprovided by computer simulations of blood flowprovides only a portion of the data which areneeded to design a treatment plan. Numerous otherfactors, including the morbidity and mortality ofthe procedure, must be considered. For example, aphysician might choose a less invasive procedureeven though it leads to less favorable blood flowthan an open procedure. The utilization of predic-tive methods will enable a physician to make thischoice with knowledge of the tradeoffs whichwould likely occur.

The predictive medicine paradigm, whereby aphysician would use diagnostic data to reconstructa model of an individual’s anatomy and physiol-ogy, and then use simulation techniques, imple-mented in a simulation-based medical planningsoftware system, will have important applicationsin medicine of the future. This approach couldallow physicians to “design” improved, patient-specific treatment plans. It is instructive to recon-


sider the “doorway to the future” described bySatava and Jones in light of these recent develop-ments in predictive medicine.41 They describe afuture scenario whereby diagnostic anatomic andphysiologic data are obtained as a patient passesthrough a physician’s doorway and a “medical av-atar” is created to help with patient education,preoperative planning, surgical simulation, andpost-operative follow-up. With the predictive med-icine paradigm proposed herein, a physician wouldnot only “simulate” a procedure, but actually im-plement and quantitatively evaluate alternate treat-ment plans using simulation-based medical plan-ning software systems like those described in thispaper. This ability to predict outcomes of alternateinterventions could emerge as a critical componentof medicine of the future.

CONCLUSIONS

The clinical application of predictive, simulation-based medical planning techniques is a fundamen-tally new approach to treatment planning. In thecase of cardiovascular disease, these methods couldenable physicians to “design” patient-specific treat-ment plans that improve blood flow. Presumably,with knowledge of the effect of hemodynamic con-ditions on vascular adaptation and disease, and thedetermination of the optimal hemodynamic condi-tions for an individual patient, better treatments canbe devised to improve patient care.

ACKNOWLEDGMENTS

The authors gratefully acknowledge the supportand assistance of Mimi Celis, Beau Vrolyk and DanWoods at Silicon Graphics, Inc., and Ben Wangand Steve Rifai at Centric Engineering Systems,Inc. The authors also acknowledge the assistance ofProfessor Thomas J.R. Hughes, Professor SandyNapel, Professor Robert J. Herfkens, Dr. RogerShifrin, and David Paik from Stanford University.This research was supported in part by the NationalInstitutes of Standards and Technology’s AdvancedTechnology Program through a subcontract withCentric Engineering Systems, Inc., The AyersFoundation, Abbott Laboratories, Boston Scien-tific, Cook Endovascular, Medtronic AneuRx, andW.L. Gore. Software was provided by Centric En-gineering Systems, Inc., TechnoSoft, Inc., XOXInc., Rensselaer Polytechnic Institute’s ScientificComputing Research Center, Stanford UniversityDepartment of Radiology, and Silicon Graphics,Inc. All computations were performed at SiliconGraphics, Inc., Mountain View, California.

REFERENCES

1. Alley MT, Shifrin RY, Pelc NJ, Herfkens RJ. Ultrafastcontrast-enhanced three-dimensional MR angiogra-phy: state of the art. Radiographics 1998;18:273–285.

2. Brooks A, Hughes T. Streamline Upwind Petrov-Galerkin formulation for convection dominated flowswith particular emphasis on the incompressible NavierStokes equations. Comp Meth Appl Mech Eng 1981;32:199–259.

3. Centric Engineering Systems Inc. Spectrum TheoryManual; 1994.

4. Delp SL, Bleck EE, Zajac FE, Bollini G. Biomechani-cal analysis of the Chiari pelvic osteotomy. Clin Ortho1990;254:189–198.

5. Delp SL, Kocmond JH, Stern SH. Tradeoffs betweenmotion and stability in posterior substituting kneearthroplasty design. J Biomech 1995;28:1155–1166.

6. Delp SL, Komattu A, Wixson RL. Superior displace-ment of the hip in total joint replacement: Effects ofprosthetic neck length, neck-stem angle, and antever-sion angle on the moment-generating capacity of themuscles. J Ortho Research 1994;12:860–870.

7. Delp SL, Loan P, Basdogan C, Rosen J. Surgicalsimulation: An emerging technology for training inemergency medicine. Presence 1997;6:147–159.

8. Delp SL, Loan JP, Hoy M, Zajac FE, Topp EL, RosenJM. An interactive graphics-based model of the lowerextremity to study orthopaedic surgical procedures.IEEE Trans Biomed Eng 1990;37:757–767.

9. Delp SL, Ringwelski DA, Carroll NC. Transfer of therectus femoris: Effects of transfer site on momentarms about the knee and hip. J Biomech 1994;27:1201–1211.

10. Delp SL, Satler K, Carroll NC. Preserving plantarflexion strength after surgical treatment for contractureof the triceps surae: a computer simulation study.J Ortho Research 1994;1:96–104.

11. Delp SL, Wixson RL, Komattu AV, Kocmond JH.How superior placement of the joint center in hiparthroplasty affects the abductor muscles. Clin Ortho1996;328:137–146.

12. Duncan DD, Bargeron CB, Borchardt SE, Deters OJ,Gearhart SA, Mark FF, Friedman MH. The effect ofcompliance on wall shear in casts of a human aorticbifurcation. J Biomech Eng 1990;112:183–188.

13. Fishman EK, Kuszyk BS, Heath DG, Gao L, Cabral B.Surgical planning for liver resection. Computer 1996;29(1):64–72.

14. Fredrickson JO, Irarrazabal P, Pelc NJ. Quantitative3D time-resolved phase contrast MR imaging. SMRWorkshop: Cardiovascular MRI: Present and Future;1994. p 25.

15. Friedman MH. A biologically plausible model ofthickening of arterial intima under shear. Arterioscle-rosis 1989;9:511–522.

16. Habbema JDF, Steyerberg EW. Predictive tools forclinical decision support. In: van Bemmel JH, Musen


MA, editors: Handbook of Medical Informatics.Houten: Bohn Stafleu Van Loghum; 1997. p 293–305.

17. Harreld MR, Valentino DJ, Duckwiler GR, LufkinRB, Karplus W. Virtual reality in radiology: virtualintervention. Proceedings of SPIE—The InternationalSociety for Optical Engineering 1995;2431:388–396.

18. Hughes T. The Finite Element Method. New Jersey:Prentice-Hall; 1987.

19. Hughes T, Franca L, Hulbert G. A new finite elementmethod for computational fluid dynamics: VIII: TheGalerkin/least squares method for advective-diffusiveequations. Comp Meth Appl Mech Eng 1989;73:173–189.

20. Jorfeldt J, Wahren J. Leg blood flow during exercise inman. Clin Sci 1971;41:459–473.

21. Kikinis R, Gleason PL, Moriarty TM, Moore MR,Alexander E, Stieg PE, Matsumae M, Lorensen WE,Cline HE, Black PM, Jolesz FA. Computer-assistedinteractive three-dimensional planning for neurosurgi-cal procedures. Neurosurgery 1996;38:640–649.

22. Krishnamurthy V, Levoy M. Fitting smooth surfacesto dense polygon meshes. In: Computer Graphics,SIGGRAPH; 1996.

23. Ku D, Giddens D, Zarins C, Glagov S. Pulsatile flowand atherosclerosis in the human carotid bifurcation:positive correlation between plaque location and lowoscillating shear stress. Arteriosclerosis 1985;5:293–302.

24. Lei M, Archie J, Kleinstreuer C. Computational de-sign of a bypass graft that minimizes wall shear stressgradients in the region of the distal anastomosis. JVasc Surg 1997;25:637–646.

25. Lei M, Kleinstreuer C, Truskey G. Numerical inves-tigations and prediction of atherogenic sites in branch-ing arteries. J Biomech Eng 1995;117:350–357.

26. Lorensen WE, Cline HE. Marching Cubes: A highresolution 3D surface construction algorithm. Com-puter Graphics 1987;21(4):163–169.

27. Lorensen WE, Jolesz FA, Kikinis R. The explorationof cross-sectional data with a virtual endoscope. Pro-ceedings of Medicine Meets Virtual Reality III, SanDiego, January 1995. p 1–10.

28. Loth F. Velocity and wall shear stress measurementsinside a vascular graft model under steady and pulsa-tile flow conditions. Georgia Institute of Technology,Atlanta, GA; 1993.

29. Nakajima S, Atsumi H, Bhalerao A, Jolesz F, KikinisR, Yoshimine T, Moriarty T, Stieg P. Computer-as-sisted surgical planning for cerebrovascular neurosur-gery. Neurosurgery 1997;41:403–410.

30. Napel S, Marks MP, Song SM, McDonnell CH, En-zmann DR, Jeffrey RB. Three-dimensional recon-structions in the evaluation of intracranial and ex-tracranial circulation using spiral CT. AmericanSociety of Neuroradiology 30th Annual Meeting, St.Louis, MO, 1992.

31. Parker D, Taylor CA, Hughes TJR, Zarins CK. Visu-alization of anatomic and physiologic data for com-

putational investigations of human vascular disease.Proceedings of the 1997 ASME Summer Bioengineer-ing Meeting, Sun River, OR, 1997.

32. Parker D, Taylor CA, Wang KC, Image based 3Dsolid model construction of human arteries for bloodflow simulations. Proceedings of International Confer-ence of the IEEE Engineering in Medicine and Biol-ogy Society, Hong Kong, 1998.

33. Perktold K, Peter R, Resch M, Langs G. Pulsatilenon-Newtonian flow in three-dimensional carotid bi-furcation models: a numerical study of flow phenom-ena under different bifurcation angles. J Biomed Eng1991;13:507–515.

34. Perktold K, Rappitsch G. Computer simulation oflocal blood flow and vessel mechanics in a compliantcarotid artery bifurcation model. J Biomech 1995;28:845–856.

35. Perktold K, Rappitsch G, Low M, Friedman MH,Kuban BD. Computer simulation of pulsatile flow inan anatomically realistic human left coronary arterybifurcation model. Adv Bioeng 1995;31:193–194.

36. Perktold K, Resch M, Peter R. Three-dimensionalnumerical analysis of pulsatile flow and wall shearstress in the carotid artery bifurcation. J Biomech1991;24:409–420.

37. Rubin GD, Dake MD, Napel SA, McDonnell CH,Jeffrey RB. Three-dimensional spiral CT angiographyof the abdomen: initial clinical experience. Radiology1993;186:147–152.

38. Rubin GD, Napel S, Dake MD, Jeffrey RB. Volumerendered CT “fly-through”—a new way to visualizeCT data. In: Society of Computed Body Tomography;1995.

39. Rubin GD, Napel S, Leung AN. Volumetric analysisof volumetric data: achieving a paradigm shift. Radi-ology 1996;200:312–317.

40. Rubin GD, Walker PJ, Dake MD, Napel S, JeffreyRB, McDonnell CH, Mitchell RS, Miller DC. Three-dimensional spiral computed tomographic angiogra-phy: An alternative imaging modality for the abdom-inal aorta and its branches. J Vasc Surg 1993;18:656–665.

41. Satava RM, Jones SB. Current and future applicationsof virtual reality for medicine. Proceedings of theIEEE 1998;86:484–489.

42. Shahidi R, Tombropoulos R, Grzeszczuk R. Clinicalapplications of three-dimensional rendering of medi-cal data sets. Proceedings of the IEEE 1998;86:555–568.

43. Shephard M, Georges M. Automatic three-dimen-sional mesh generation by the finite octree technique.Int J Num Meth Eng 1991;32:709–749.

44. Steinman, DA, Ethier CR. The effect of wall disten-sibility on flow in a two-dimensional end-to-side anas-tomosis. J Biomech Eng 1994;116:294–301.

45. Taylor C, Hughes T, Zarins C. Computational inves-tigations in vascular disease. Comp Phys 1996;10:224–232.


46. Taylor C, Hughes T, Zarins C. Finite element model-ing of blood flow in arteries. Comp Meth Appl MechEng 1998;158:155–196.

47. Taylor CA, Hughes TJR, Zarins CK. Finite elementmodeling of 3-dimensional pulsatile flow in the ab-dominal aorta: relevance to atherosclerosis. AnnalsBiomed Eng 1998;26: 975–987.

48. Taylor CA, Hughes TJR, Zarins CK. Effect of exer-cise on hemodynamic conditions in the abdominalaorta. J Vasc Surg 1999;29:1077–1089.

49. Wang K, Taylor CA, Hsiau Z, Parker D, Dutton RW.Level set methods and MR image segmentation forgeometric modeling in computational hemodynamics.In: 20th Annual International Conference of the IEEEEngineering in Medicine and Biology Society, HongKong, 1998.

50. Womersley J. Method for the calculation of velocity,rate of flow and viscous drag in arteries when the pres-sure gradient is known. J Physiol 1955;127:553–563.

51. XOX. SHAPES Reference Manual; 1994.52. Young GS, Silverman SG, Kettenbach J, Hata N,

Golland P, Jolesz FA, Loughlin KR, Kikinis R. Three-dimensional computed tomography for planning uro-logic surgery. Urologic Clinics of North America1998;25(1):103–111.

53. Zarins CK, Giddens DP, Bharadvaj BK, Sottiurai VS,Mabon RF, Glagov S. Carotid bifurcation atheroscle-rosis: Quantitative correlation of plaque localizationwith flow velocity profiles and wall shear stress. CircRes 1983;53:502–514.

54. Zarins CK, Taylor CA. Hemodynamic factors in ath-erosclerosis. In: Moore W, editor: Vascular Surgery:A Comprehensive Review. Philadelphia: W.B. Saun-ders;1998. p 97–110.

55. Zarins CK, Zatina MA, Giddens DP, Ku DN, GlagovS. Shear stress regulation of artery lumen diameter inexperimental atherogenesis. J Vasc Surg 1987;5:413–420.


Predictive medicine: Computational techniques in...

Documents

Transcript of Predictive medicine: Computational techniques in...