This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

**Rockwell Mackie SIAM Review, Vol. 41, No. 4. (Dec., 1999), pp. 721-744.
**

Stable URL: http://links.jstor.org/sici?sici=0036-1445%28199912%2941%3A4%3C721%3AOTDORT%3E2.0.CO%3B2-G SIAM Review is currently published by Society for Industrial and Applied Mathematics.

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/journals/siam.html. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission.

The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academic journals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers, and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community take advantage of advances in technology. For more information regarding JSTOR, please contact support@jstor.org.

http://www.jstor.org Wed Dec 12 03:11:59 2007

SlAM REVIEW Vol. 41, N o 4. pp. 721-744

@ 1999 Society for Industrial and Appl~edMathemat~cs

**Optimizing the Delivery of Radiation Therapy to Cancer Patients*
**

David M. Shepard' Michael C. ~ e r r i s t Gustavo H. oliverat Rockwell Mackiei

Abstract. In the field of radiation therapy, much of the research is aimed at developing new and

innovative techniques for treating cancer patients with radiation. In recent years, new treatment machines have been developed that provide a much greater degree of computer control than was available with the machines of previous generations. One innovation has been the development of an approach called "tomotherapy." Tornotherapy can be defined as computer-controlled rotational radiotherapy delivered using a n intensity-modulated fan beam of radiation. The successful implementation of the new delivery techniques requires the development of a suitable approach for optimizing each patient's treatment plan. One of the challenges is t o quantify optimality in radiation therapy. We have tested a variety of objective functions and constraints in pursuit of a forinulation that performs well for a wide variety of disease sites. An additional challenge stems from the sizable amount of data and the large number of variables that are involved in each optimization. This paper presents several approaches to optimizing treatment plans in radiation therapy, and the advantages and disadvantages of a number of formulations are explored.

Key words. optimization, radiation therapy, mathematical programming, cancer, treatment plans AMS subject classifications. Primary, 90C50; Secondary, 9 0 C l l . 90C30 PII. SO036144598342032

I . Introduction. This year in the United States more than 1,200,000 people will be diagnosed with cancer [55]. More than half of these cancer patients will be treated with radiation at some point during the course of their disease [66]. Radiotherapy refers to the use of radiation as a means for treating disease. A beam of radiation is attenuated as it passes through a patient, and energy is deposited at each location along the path of the beam. The patient's dose is the

"Received by the editors July 20. 1998; accepted for publication (in revised form) April 13, 1999; published electronically October 20, 1999. This material is based on research supported by Air Force Office of Scientific Research grant F49620-98-1-0417, National Science Foundation grant CCR9619765. and National Institute of Health grant CA 48902.

http://www.siam.org/journals/sirev/41-4/34203.l~tml

t ~ e ~ a r t m e of t Medical Physics, University of Wisconsin-Madison, 1300 University Ave., Rm. n 1530, Madison, W I 53706 (shepardOmadrad.radiology.wisc.edu, oliveraOmadrad.radiology.wisc.edu, mackieOmadrad.radiology.wisc.edu, http://www.madrad.radiology.wisc.edu). The fourth author has a financial interest in TomoTherapy Incorporated, which may participate in the commercialization of tomotherapy. $Computer Sciences Department. University of Wisconsin-Madison, 1210 West Dayton Street; Madison, W I 53706 (ferrisOcs.wisc.edu).

The physician must also consider the patient's medical history and the institution's capabilities. and the optimizations were performed for a. each of which is assigned a separate beam of intensity. tomotherapy. Thus. physicians develop insights into what is "optimal" for a particular type of cancer in a particular location. because the repair mechanism of cancer cells is less efficient than that of normal cells. is described in detail. was chosen because of its strength as a tool for efficiently performing largescale optimizations. MATLAB [48] and GALIS 191. the General Algebraic Modeling System. single transverse cross section of the patient. we have tested a variety of mathematical programinii~gapproaches to this problem. It is hoped that this research will provide insights to both developers of optimization technology and physicians about the potential capabilities of a carefully devised optimization approach.1. it is iinportaiit that the optimization approach provide the physician with sufficient flexibility so that he or she can always produce an acceptable treatment plan. advances in radiation therapy have led to the development of delivery techniques with a high degree of computer control. LIATLAB was chosen because of its capabilities in both nuineric computation and advanced graphic visualization. Section 4 of this paper presents the results of an investigation into a variety of optimization approaches for determining each patient's treatment plan. One linear programming approach is to minimize the total dose delivered to the patient subject to a lower bound on the . One of these techniques. This is true of both cancer cells and norinal cells.722 D. Unfortunately. our initial investigation into a variety of optimization techniques did not require that level of complexity. In recent years. the goal is to provide the highest probability of tuinor control while maintaining an acceptable probability of radiationinduced complications. we have used precomputed photon beamlets as a means of testing a 1 large number of objective functions and constraiilts [GO. M. G. These delivery techniques provide the user with a large number of **beamletsn radiation. FERRIS. GAMS. These techniques offer many new opportunities in the delivery of radiation therapy. however. If all of the tuinor cells are killed. AND T. Radiotherapy is a useful tool for treating cancer. OLIVERA. Throughout this work. When designing a patient's treatment plan. High doses of radiation can kill cells or prevent thein from growing and dividing. Linear programming forinulations are discussed in section 4. Due to the complexity of the treatments. This environment provided a means for efficiently testing models that varied widely in terms of both the complexity of formulation and the time required to obtain an optimal solution. 7 . This simple setup was used in order to speed up the dose computations [70]. and the quality of the solutions that are produced using each of these approaches. This project made use of two inathenlatical software packages. Section 3 introduces the more advanced delivery techniques that have been developed in recent years. an optimization approach is needed in order to develop the best plan for treating each patient. Thus. MACKIE amount of energy deposited per unit mass. the patient is cured. normal cells are more likely to fully recover from the effects of radiation. Over years of practice. This paper outlines the advantages. it is very difficult to quantify optimality in radiation therapy. A simulated "patient" was represented by an all-water cylinder. Section 2 of this paper provides an overview of radiation therapy and conventional delivery techniques. The treatments are carefully planned so that the radiation damage is concentrated within the cancerous region [62]. In pursuit of a suitable approach for optimizing each patient's treatment plan. the disadvantages. SHEPARD. Although actual patient data can be used with our algorithms.

particularly when modeling tools and optimization packages are used. Section 5 draws some conclusions and outlines research challenges in this area. The highest dose level appears at the top of the picture. 2.3 presents two implementations of partial volume constraints. beams are delivered from a number of different angles spaced around the patient. because the beam enters from the top. To avoid this. One implementation is based on the use of a nonlinear ramp function. Nonlinear optimization approaches are described in section 4. The patient lies on the couch of the treatment unit. we would also risk severe damage to any normal tissues located within the path of the beam of radiation. The x-rays are then shaped into a beam of radiation that is directed toward the patient. dose to the tumor and an upper bound on the mean dose to the region at risk. This is a result of the fact that the beams intersect within the tumor. This figure can be thought of as a simplistic representation of a transverse cross section through a patient.3). This provides the ability to treat the patient from a wide range of beam directions. In Figure 2.4. Section 4. The vast majority of patients who receive radiation are treated on a machine called a linear accelerator (see Figure 2. Note that the dose to the tumor is significantly higher than the dose to the surrounding normal tissues. and the second is based on mixed integer programming (MIP).1 The dose distribution from a single beam of mdiation.1 presents the distribution of dose from a single beam of radiation. If we were to treat a patient with a single beam of radiation. The inner circle denotes the outline of the tumor. . Extensions of our models to conformal avoidance problems are provided in section 4. Figure 2. The outer circle represents the outline of the patient. Linear programming formulations such as this can be solved very efficiently. it shows the dose to the central slice of a 20-cm-diameter all-water cylinder irradiated with 2-MeV x-ray photons. 2. a dose distribution from the delivery of five equispaced beams of radiation is shown. The head of the linear accelerator can be rotated around the patient. and the inner circle denotes the outline of the tumor. An appreciation of how a beam of radiation is attenuated in a human body helps one to better understand radiation therapy. it might be possible to kill all of the tumor cells.2. More specifically.2.OPTIMIZING THE DELIVERY OF RADIATIONTHERAPY - - Fig. The x-rays produced by the linear accelerator are filtered so that the beam of x-rays is uniform in intensity as it exits the machine. however. Unfortunately. These formulations are primarily based on a weighted least squares approach. Conventional Radiotherapy.

Note that the dose is wncentmted in the tumor.D. 2.3 A clinical linear accelemtor used to treat cancer patients. Radiotherapy patients are typically treated with between three and seven beams of radiation. Fig. where the beam intersect. Figure 2.2 A dose distribution created with five equispaced beams.4 shows three treatment plans developed using conventional uniform beam delivery. OUVER4. The beam angles.4(b) shows the dose distribution for a target of intermediate complexity. this . the beam intensities. The target is U-shaped with a region at risk placed in the concavity of the U. In conventional treatments. Figure 2. the dashed line represents the 90% isodose curve. This line indicates a dose level that is equal to 90% of the dose prescribed to the target. Through an iterative process of trial and error.4(a) shows the optimized dose distribution for a simple square-shaped target with a smaller square shaped region at risk located adjacent to the target. This type of geometry would be seen clinically if a tumor had grown around a patient's spinal cord or brain stem. a radiation oncologist and a physicist determine an acceptable set of beam angles and beam weights that produces an acceptable treatment plan. a uniform or linearly varying beam intensity is delivered from each beam angle. G. Ideally. FERRIS. The target is Gshaped with a region at risk located between the arms of the L. In all three cases. M.4(c) shows the dose distribution for a complex target shape. and the predicted dose distribution together make up the treatment plan. SHEPARD. MACKIE Fig.4. 2. Figure 2. seven equispaced beam angles were used. A N D T. Figure 2. In Figure 2.

Seven equiepaced beam angles were w e d . This is done in order to eliminate the possibility of radiation-induced paralysis. Because of the high dose to the region at risk.4(a) that uniform beam delivery is able to closely match the 90% isodose line with the border of the square-shaped target.5. Figure 2. however. It should be noted.4(c) reveal that uniform beam delivery does not permit the sparing of a region at risk located in the concavity of a more complex target shape. The dashed line represents the 90% ieodose curve. physicians often choose to block out any portion of a beam that passes through a patient's spinal cord. A simple method for decreasing the probability of normal tissue complications is to block any portion of a beam that passes through a region at risk.4(b) and 2. line should closely match the boundary of the target. segmented beam delivery was successful at sparing the region at risk. Figures 2. Segmented beam delivery is of particular importance in cases where it is essential to avoid complications.5 displays three treatment plans developed using segmented beam delivery. For example. This is referred to as segmented beam delivery. It can be seen in Figure 2. This lane andiccrtes a dose level equal to 90% of the prescribed dose to the target. The 90% isodose curve is shown as a dashed line in a number of figures throughout this paper. that with the intermediate . Unmodulated fields were w e d with the field size set to match to target dimensions.4 Three treatment plans developed wing uniform beam delivery. In all three cases shown in Figure 2. 1. this treatment could lead to a high probability of complications.OPTIMIZING T E D LV R O RADIATION T E A Y H EIE Y F HRP (a) Simple: Square Target (b) Intermediate: L-Shaped Target ( c ) Complex: U-Shaped Target Fig.

shape. 40. and location of a patient's tumor has been dramatically improved by the development of computed tomography (CT) and magnetic resonance imaging (MRI). 171 demonstrated that the use of nonuniform beam intensities permits the delivery of dose distributions that conform to the target shape while simultaneously sparing any neighboring regions at risks. G. In IMRT. The areas of underdosage could permit the survival of some of the cancer cells. Seven equispaced beam angles were used. much of the current research is aimed at developing better hardware that will make it possible to take full advantage of these improved imaging capabilities. Independent investigation performed by Brahme [5] and Cormack [16. OLIVERA. one can think of each beam as being divided into a large number of . 2.5 Three treatment plans developed using segmented beam delivery.D. the ability to pinpoint the size. 49. M. and this could eventually lead to a recurrence of the disease. More complex treatment plans to alleviate these problems will be described in what follows. 11. In radiotherapy. 8. (a) Simple: Square Target (b) Intermediate: L-Shaped Target (c) Complex: U-Shaped Target Fig. Open and unmodulated fields were used. FERRIS. and the portion of the field passing thmugh the region at risk was blocked out. portions of the tumor were significantly underdosed. 22. each treatment beam has an individually defined intensity distribution. Recent technological advances have made possible the delivery of intensity modulated radiation therapy (IMRT) [71. 751. SHEPARD. In other words. 4. A N D T MACKIE . 3. 26. and complex target shapes. In recent years. 3. IMRT and Tornotherapy.

Figure 3 1 shows three dose distributions produced using IMRT with seven equi. spaced beam angles. . 43. 29.The dosimetric advantage of conformal therapy can also be used as a means of providing a reduced probability of normal tissue complications [28]. smaller beamlets of radiation.60. 12.-Shaped Target (c) Complex: U-Shaped Target Fig.) 46. 30. 31. It can be seen that for all three setups. The development of IMRT represents an important step in the realization of conformal therapy. the region at risk was spared without reducing the probability of tumor control. 15. The goal of conformal therapy is the delivery of a high dose of radiation to the tumor while simultaneously sparing any neighboring regions at risk. 1.2 1 0. 2 1 A number of key features distinguish tomotherapy from conventional 3.Seven equispaced beam angles were used with a beamlet size of 6 mm. Approximately 80 beamlets were used in each of these cases. IMRT was able to successfully conform the 90% isodose line to the outline of the target shape. Successful delivery of conformal therapy should allow for an escalation of the tumor dose and may enhance local tumor control [41]. 67. 6. The weights of these beamlets can be optimized so as to produce the most favorable dose distribution [70.2. 3. 50. One approach to the delivery of IMRT is called tomotherapy (see Figure 3 2 [45.61. 77. Thus.OPTIMIZING THE DELIVERY OF RADIATION THERAPY (a) Simple: Square Target (b) Intermediate: I. .1 Three treatment plam developed using IMRT.

3. f Fig. meaning each leaf is either in the field or out of the field. Each leaf of the collimator can be w e d to block a portion of the beam.3 In the housing of the tomoth- machine..3. and it progresses around the patient at a constant speed. G. A N D T MACKIE Fig. OUVERA. The linear accelerator is mounted on a ring-shaped gantry. FERRIS. M.2 An artist's rendering o a tomothempy machine. SHEPARD.D. The collimator is binary in nature. treatments. . One characteristic of tornotherapy is the use of a rotational treatment delivery. the fan beam of mdiation posses through a multileaf collimcrtor.

The hardware in the tomotherapy machine will also provide a means of obtaining an accurate reconstruction of the dose delivered to the patient. This is accomplished through the use of a moving couch. Tomotherapy promises to provide a number of enhancements in addition to its improved beam-delivery capabilities. 651. Optimization Approaches. Through optimizing the beam weights. Over the course of the treatment the fan beam is able to treat the entire tumor volume. Radiation oncologists often use a cumulative dose volume histogram as a means of determining the quality of a treatment plan (see Figure 4. Tomotherapy does not make use of a broad beam of radiation. Originally. the dose optimizations were performed using the routines provided in MATLAB'S Optimization Toolbox. a tomotherapy machine will be able to produce C T images of the patient [69]. For example. This technique was chosen because of its ability to accurately model complex dose distributions [47. it is difficult to quantify the goals of radiation therapy. and w p are the beamlet weights. The matrix D represents the total dose delivered to a transverse cross section of the patient. Within the housing of the machine. As outlined in section 1. The amount of time that an individual leaf is open during a step is a variable that is referred to as a beam weight. the dose to the patient from each individual beamlet is computed. i and j are the pixel coordinates. the fan beam passes through a device called a multileaf collimator (see Figure 3. The collimator has a series of high-density leaves. The dose matrices were computed using the convolution/superposition technique. Another feature of tomotherapy is that it will provide precise registration capabilities designed to ensure that the location of the patient on the treatment couch is consistent from day to day [25]. A cumulative dose volume histogram displays the fraction of the patient that receives at least a specified dose level. MATLAB'S algorithms performed reasonably well . 4. any variation between the delivered and the predicted dose distributions can be compensated for in subsequent treatments [63].OPTIMIZING THE DELIVERY OF RADIATION THERAPY 729 Tomotherapy is also distinguished by the size of the beam that is used to treat the patient. and each leaf can be used to block out a portion of the fan beam. Due to the large number of beams and the range of intensities involved. and it equals the sum over all of the beamlets p of the beam weight multiplied by its corresponding dose matrix DP: where n is the number of beamlets.1). Each full rotation of the linear accelerator can be approximated as a series of discrete steps. Instead. The patient lies on the couch of the tomotherapy unit and is slowly advanced through the rotating fan beam of radiation.3). it is possible to deliver a dose distribution that closely conforms to the target shape. As the linear accelerator progresses around the patient. IMRT treats each patient with thousands or even tens of thousands of beamlets of radiation. The first step in the development of a treatment plan is to define the location of the tumor and the regions at risk (critical structures). The remainder of this paper is devoted to investigating this issue. the leaves move into and out of the beam. Thus. Next. a computer-based optimization routine is required in order to develop each patient's optimal treatment plan. the beam is collimated into a fan beam that delivers radiation to a narrow slice of the patient.

and automatic differentiation far outweigh the slight overhead imposed by a modeling language. The advantages and disadvantages of a number of approaches were tested and the results are described in the following subsections. and mixed integer models. Selecting or changing the problem formulatioils is very simple in a modeling language environment. modification. and maximum doses to each region. Another concern was the fact that physicians may need to make a number of changes in the parameters of an optimization before they are satisfied with the treatment plan.1 A dose volume histogram from an optimized treatment plan. the dose distribution for each beamlet was stored in a relatively dense 100 by 100 matrix. The development of an optimal treatment plan is made particularly challenging as a result of the large number of variables and the sizable amount of data that are involved in each optimization. For our simulated treatments. minimum. Many of the figures shown in this paper were produced using standard MATLAB graphics utilities along with the optimization results generated using GAMS. t\ % 1. 4. SHEPARD.2 0.4- % L E 0. however. . We thus turned to GAMS in hopes of increasing our optimization capabilities. CONOPT [21].8.g 0. All of these solvers are very sophisticated large-scale implementations of state-of-the-art algorithms for their respective problem classes. GAMS also provides the ability to switch between a number of commercial solvers including OSL [32].8 - '. * * I * . A N D T MACKIE . 5 I 0. 1. '. As the number of beamlets increased. OLIVERA. M. The standard deviation in dose over the target was also determined and the dose values were displayed using a histogram. At the end of each simulated treatment. Therefore. the speed of each technique is an important concern. the problems became too large for MATLAB to solve in a reasonable amount of time. 1 \ 1 \ ! ! LI 'I I .4 0. Up to 7.69 0 \ r! I '1 + E 0 .2 1. MATLAB was then used to visualize and analyze the results. Currently.000 beainlets were used in an optimization. The advantages of easy problem formulation. nonlinear. I I . ' . I * 0 0 0. CPLEX [33]. We believe that only after a particular optimization approach becomes prevalent would it make sense to use one of these solvers exclusively. and MINOS [56].0. G.D. each simulated treatment is set up within MATLAB and a simple interface [24] is used to pass the data to GAMS. FERRIS.6 Relative Dose -*. for small-scale optimization problems. The user can choose between linear. the optimized beam weights were returned from GAMS to MATLAB using the aforementioned interface.4 1 Fig.'\ Target b \ 5 ' 00. The analysis of the results included a determination of the mean.21 ! i.

"=. n is the number of beamlets. The objective function can be modified so as t o minimize a weighted integral dose: R is the subset of pixels in the region at risk. QT is the target weight. j). The linear programming problem is of the form min subject t o cTx Ax 2 b. Linear Programming Approaches. Y < Dk1 q k . ~ < x < u . Q R is the region at risk weight. this can be reformulated as follows: > min. There are a number of potential linear programming approaches t o radiotherapy dose optimization [68]. Another linear programming approach is t o place both an upper and a lower bound on the dose t o the tumor. and e and u are vectors of lower and upper bounds on the variables. The most widely used mathematical programming formulation is undoubtedly the linear program [13.1.j) Dij Dij = c. 18. By increasing the relative weight assigned t o a region at risk.OPTIMIZING THE DELIVERY OF RADIATION THERAPY 73 1 4. the user can place a greater emphasis upon achieving a uniform target dose. wPDFj 'j(i. using either sophisticated implementations of the simplex method [13. By increasing the relative weight assigned to the tumor. 59. In this formulation. the objective might be t o minimize a . cTx describes the linear objective function. Linear programming codes have successfully solved many large-scale optimization problems. 1) 7. 18. and y is the lower bound on the dose t o the tumor.1). Ax b represents the set of constraints. subject to C(i. where x is the vector of continuous decision variables. A nonnegativity constraint is placed on the beam weights. the user can place a greater emphasis upon reducing the dose t o that region. where 7 is the subset of the pixels located in the tumor. In this case. The integral dose is the total dose summed over all of the pixels. 641 or adaptations of primal-dual interior point methods [76]. The results from a simulation using this approach are shown in Figure 4. and QN is the normal tissue weight. 571.2(a). a weighting factor is assigned t o each region of the patient. The objective function equals the sum over the entire volume of each pixel's dose multiplied by its weighting factor. The above formulas are designed so that the optimizer will drive the tumor dose down t o the specified lower bound. One linear programming formulation is t o minimize the total integral dose subject t o a lower bound on the dose t o the tumor. and N is the subset of normal tissue pixels (those located outside both the target and the sensitive structures). Using the notation from (4.

(d) The maximum deviation in dose over the tumor was minimized subject to the constraint that the mean dose to the region at risk had to be kept below 0. (c) The dose is more evenly distributed outside the tumor as a reault of a constraint that the maximum beam weight could not be more than a factor of 5 greater than the mean beam weight.9.l)eR + Dij = C" w P DP j p =l s v(4 j ) .2 A series of four dose distributions p m i d using linear progrcrmming techniques. Figure4.D. V(k. Fig. The U-shaped taryet was cut out of a 5 cm by 5 cm square. SHEPARD. When minimixing the integral dose. FERRIS.1 and a lower bound of 0. 15 h m angles were wed along with a collimator size of 6 mm. G. In all cases. OUVERA. (b) An upper bound of 1. 1) E 7.2(b) displays a dose distribution produced using this approach with ON equal to 1 and OR equal to 2. weighted integral dose swumed over al of the nontumor pixels: l mi% ON C ( ~ .9 were placed on the target dose.5. ~ ) E N C(k. AND T. MACKIE (a) The optimizer sought t o minimize the integral dose subject t o a lower bound on the tumor dose of 0. subject to TL 5 Dkl 5 TU TL and yu indicatethe lower and upper bounds on the dose to the target. The objective function was a weighted integral dose. the dashed line denotes the 90% bodme curve. The . the highest weights are assigned to the beamlets that deposit the greatest fraction of their integral dose within the tumor. M. 4.

1) I D M < YU C ( r ..< :CF=l~P.An upper limit of /3 was placed on the mean dose t o the region at risk.2(c). Overall. 4. for example. 1) 6 7 . the maximum beam weight was constrained to be less than a times the mean beam weight. see [13]. ~ I x < u . I DMI I wp YU WP. One can only devise a relatively limited number of objective functions and constraints that fall within the linear realm. This can be accomplished using the following model: min. A solution to this problem is to place an upper bound on the ratio between the maximum beam weight and the average beam weight: min. subject to one or more constraints. subject to C(~. Y(k. The two primary advantages of the use of a linear programming approach to treatment planning optimization are its speed and the ease of formulation. 7. s ) c RDTS I ~ R P .2(d) shows an optimized dose distribution produced with this model. . w. In Figure 4.2(c) appears to be the best linear programming result for this particular target shape. A concise mathematical description of the nonlinear programming (NLP) problem is as follows: min subject to f (x) g(x) 0. n .6ij 1 ~~ Dij = C:=l YL w. In the above equation.~)ER + 02. it is unlikely that a physician could always achieve an acceptable result.. 6. the dose distribution in Figure 4. subject to m a x ( i . .D$ 'd(4 j ) . In this case. the maximum beam weight could not exceed a times the mean beam weight. 2. A disadvantage of the linear programming formulations is a lack of flexibility. . 2. -p=l YL Y(i1j)) Y(k. A final linear programming approach is to minimize the maximum deviation from the prescribed target dose.=l > 0. m = 1 . Once again.n. : C. < . Heavily weighted beams produce streaks of high dose through the patient. n n is the number of pixels in R. it can be seen that this approach resulted in a more evenly distributed integral dose as compared to the previous linear programming optimizations.OPTIMIZING THE DELIVERY OF RADIATION THERAPY 733 results can become unsatisfactory if any of the beam weights is made exceedingly large. m = 1 . j )lDij .2. Figure 4. . and this could lead to patient complications. Nonlinear Programming Approaches. En wpD$ .~)ENQ C(~. With any particular linear programming formulation. w . The technique used to reformulate a minimax problem as a linear program has been well documented.. .7 .

the optimizer sought to minimize the weighted squared differences between the prescribed and the actual doses summed over all of the pixels. Unfortunately. G. A drawback to this approach (and those outlined in the previous paragraph) is that increasing the complexity of the formulation will often lead to more time-consuming optimizations. With a nonlinear formulation [27]. Another weakness of the NLP approach is that for general functions. Thus. MACKIE Here. one may need to run a series of optimizations. each weighting factor was normalized by the number of pixels in the region to which the weight was assigned. For a particular patient. For example. 44.. From our investigations into NLP approaches. This problem is a bound-constrained weighted least squares problem and can therefore be solved by various specialized large-scale optimization algorithms [14. 10. The values of Dij(w) are determined using (4. in order to obtain an acceptable result. This is an area of future research. In both cases. 37. we have chosen not to include these models in our current study. The ability to predict the probability of normal tissue complications is particularly suspect. 361. FERRIS. Ideally. 51. 52. This helps to eliminate difficulties with a large structure dominating the optimization.1 give rise t o general constrained nonlinear programs. The importance of multiple . we have used a weighted least squares objective function. an oncologist could choose to maximize the probability of tumor control subject to a cap placed upon the probability of complications for each sensitive structure. Formulas have also been designed in hopes of predicting the probability of normal tissue complications [38. Because of the degree of uncertainty involved. x is a vector of variables that are continuous real numbers. AND T. and g(x) represents the set of constraints. 341. f (x) is the objective function. the target was prescribed a dose of 1. It should also be mentioned that nonlinear formulations have been used by a number of researchers in an attempt to use mathematical models as a means of predicting biological response. For many of our simulations. this type of biological modeling could serve as a very useful tool in radiotherapy optimization.734 D. the optimizer can only guarantee that the solution is locally optimal. The objective function is As before. 42. The matrix S describes the prescribed dose. and ON is the normal tissue weight. QRis the region at risk weight. S is typically set equal to zero. we have concluded that the primary advantage of using this approach is that it is possible to devise numerous complicated objective functions and constraints. For our simulations. t and u are vectors of lower and upper bounds placed on the variables. For example. M. Outside the target. QT is the target weight.3 displays two optimized dose distributions produced using different weighting schemes. SHEPARD. the best choice of weighting factors is not always intuitive. the current biological models require input parameters that are not known with great certainty. researchers have devised mathematical formulas that are designed to predict the probability of tumor control [72. 581. In these cases. Ratio constraints on average and maximum beam intensities as well as other bounding constraints outlined in section 4. OLIVER. there is an expanded range of possible objective functions and constraints as compared with linear programming.1). Figure 4.

the radiation oncologist is willing to sacrifice a portion of a region at risk in order to improve the probability of curing the disease. (b) The target.3 The results fmm two simulations are shown. region at risk. 4. Fig.1) displays the fraction of each region of the patient that receives at least a specified dose level.541.OPTIMIZING THE DELIVERY OF RADIATION THERAPY (a) The target. We believe that it is possible to develop convex objective functions that adequately describe the problem. (d) Cumulative dose volume histogram for (b). 5. a locally optimal solution is also globally optimal.60]. (c) Cumulative dose volume histogram for (a). respectively. and normal tissues were assigned weights of 10. In both m e s .1. 4. Partial Volume Constraints. region at risk. local minima in radiotherapy optimization has been investigated elsewhere [19.3. Thus. and 2. Oncologists often specify constraints of the form "No more than x% of this refion . and 1. The cumulative dose volume histogrum plots the fraction of each structure that receives at least the dose specified by the abscissa. It is our hope that a deterministic method will provide a more rapid approach for developing a patient's treatment plan. in this case. and normal tissue were assigned weights of 25. a number of researchers in radiotherapy optimization have made use of simulated annealing approaches (74. In some cases. h a l l that the dose volume histogram (see Figure 4. Note that the least squares formulation presented above was convex. Note that the tumor dose is more unafonn when the tumor was assigned a higher relative weight. Because of the potential to become stuck in a local minimum. 73.35. a weighted quadratic objective function was used.

the factor of 3 in the denominator creates a more sloped ramp. In this case the error function (erf) was used (see Figure 4. The nonlinear formulation represents a new approach to partial volume constraints: min. f (x) = erf(x). For this simulated treatment. Figure 4. MACKIE (a) The error function used in our implementation.5 shows the results of an optimization performed using this approach. The . the dose limit will be denoted by A and the fraction of the volume allowed to exceed this limit will be denoted by R. the arms of an L embrace a smaller.j).4(a)). For each partial volume constraint. FERRIS. This type of requirement is called a partial volume constraint [53.) 2 ( D23 . SHEPARD. Fig. For our formulas. The goal of the optimizer was to minimize the sum of the squared difference between the prescribed and the actual doses over all of the pixels in the tumor subject to two partial volume constraints. n~ is the number of pixels in the region at risk and nN is the number of pixels in the normal tissue. We have tested both a nonlinear and a mixed integer approach t o the implementation of partial volume constraints. .- zn w pDfj p=l v(i. Akl) 5 nRfl'R1 C(k.D. M. f (x) = erf(xl3). at risk can exceed a dose of y.23. In these formulas. square-shaped target. 4.j)E~ 6. for a particular region at risk. the oncologist determines both a dose limit and a fraction of the structure that can exceed the dose limit. the error function was shifted so that the center of the ramp matched the dose limit.4 E r r o r functions used for partial v o l u m e constraints. 391.l)ER erf(Dk1 - 023 . The geometry is the inverse of the L-shaped target used previously. OLIVERA. > 0. . This geometry was chosen in order to effectively test this implementation of partial volume constraints. the prescription dose level in the tumor was set at 70 Gy." Thus.l)EN erf(Dk~ - AM) IN R N . . (b) A modified error function (erf(xl3)). In this case. With other shifts and scalings it may be possible to improve the solution quality and numerical performance. AND T. The partial volume constraints were realized through the use of a ramp function. G. subject to z(i. z(k. ~ w.

27 w p > 0.5 Optimization using partial volume constmints implemented via a m m p function. 1. It may be advisable to first solve a problem using a ramp function with a gradual slope. One must be careful. The goal of this optimization was to minimize the maximum deviation + > . A scalar value can be placed in the denominator of the error function in order to alter the slope of the curve (see Figure 4.OPTIMIZING THE DELIVERY OF RADIATION THERAPY ! 0 1 0 2 0 3 0 4 0 5 0 W 7 0 8 0 Relative Dose ( a ) optimized dose distribution ( b ) cumulative dose volume histogram Fig. If the slope of the ramp is too gradual. x is a vector of variables that are continuous real numbers and y is a vector of variables that can only take integer values. first partial volume constraint specified that 80% of the region at risk must be kept below 30 Gy. . That solution can then be used as an initial guess in an optimization using a steeper ramp function. The diamond on the dose volume histogmm indicates the dose volume histogmm wnstmint applied to the critical structure. The results presented in Figure 4. MIP makes possible a second approach to the implementation of partial volume constraints.4(a). ~ S X ~ U . Yi In this formulation. and Ax B y lower and upper bounds placed on the continuous variables. and the triangle indicates the dose volume histogmm constmint applied t o the normal tissue. The second partial volume constraint specified that 95% of the normal tissue region must be kept below 50 Gy. 20% of the region at risk was allowed to exceed 30 G and 5% of the region at risk was allowed to exceed 50 Gy. however. . and y E {O. On the other hand. Both of these possibilities occurred in our experimentation. As the ramp function is made steeper. An optimized dose distribution produced using a mixed integer model is shown in Figure 4. C and u are vectors of function. a ramp that is too steep may prove to be too difficult for the optimization algorithm.5 were produced using the error function depicted in Figure 4. + B y > b. ) is the integrality requirement. the MIP problem appears as follows: min subject to cTx Ax +dry }. the constraint becomes more absolute. Some related work using MIP formulations is described in [39]. 4. In this y case. the solution may differ considerably from the desired results.. 2.4(b)). cTx+dTy is the linear objective b represents the set of constraints. when choosing the scaling factor. {O. Mathematically.1.6.

4. SHEPARD. With the ramp function technique. This.D. the optimizer summed Or. The following formulation was used to generate the simulated treatment: min. here 15% of the critical structure was allowed to exceed 30 Gy. however. any linear or nonlinear objective function or constraint can be used. A limitation of the MIP implementation of partial volume constraints is that only linear objective functions and linear constraints can be included in the optimization. subject to The constraint is a standard MIP technique for modeling an "if-then" constraint.6 An MIP implementation of partial volume constmints. in dose over the target subject to a partial volume constraint that specified that 85% (aR) of the region at risk must be maintained below the dose limit (A) of 30 Gy. so we typically set M equal to 100. OUMRA. then Okl must be set equal to 1. The variables Okl can then be used to enforce the partial volume constraint. AND T. over all of the pixels in the region at risk. MACKIE (a) optimized dose distribution (b) dose volume histogram Fig. FERRIS. the dose at any pixel is always less than 100 Gy. can lead to more complex and difficult nonlinear p r e grams that increase the probability of finding local solutions that are not globally . One appealing quality of partial volume constraints is that physicians are experienced with this type of requirement. G. and it required that this value be less than OR times the total number of pixels in the region at risk. Partial volume constraints are much more intuitive than assigning relative weights to each region of the patient. Note that if Dkl exceeds Akl. In our case. In the above case. The value of M is chosen to be large enough so that the corresponding constraint is trivially satisfied when Okl = 1. This constraint was modeled through the use of a binary variable Oij that equaled 1 if the pixel in the region at risk exceeded the prescribed limit and 0 otherwise. M. however.

A large nodal region was prescribed a relative dose of 5. Conformal Avoidance.7 and 4. due to the complex formulation. This case utilized a U-shaped target with a prescribed relative dose of 7. Instead. Note that a large number of beam angles are needed to approach this dzflcult prescription. there is little rationale for attempting to create a dose distribution that conforms tightly to the tumor volume.7 A series of confonnal avoidance dose distributiolrs. The goal of the optimization was to produce a treatment plan that provided a uniform dose over both the target and the nodal regions. the optimizer attempted both to conformally avoid a sensitive structure in the concavity of the U and to minimize the dose to the nonnodal region. 4. Thus. The primary disadvantage of partial volume constraints is that. We have studied the possibility of using intensity-modulated beams in order to deliver conformal avoidance radiotherapy.OPTIMIZING THE DELIVERY OF RADIATION THERAPY (a) prescription (b) 7 beam angles ( c ) 15 beam angles (d) 71 beam angles Fig. it may be beneficial to design a treatment that treats a regional field while "conformally avoiding" any critical structures in that field.8. This blurring makes it impossible to . optimal. 4.4. For a large number of radiotherapy patients. The results from a series of simulations are shown in Figures 4. each constraint can add considerable time to the optimization. Simultaneously. the tumor volume is not clearly defined. It should be noted that finite source size and the phenomenon of electron transport both cause a blurring of each bearnlet.

.\ . i -.-$ . Conclusions and Future Work. G. One method to speed up conformal avoidance optimizations is to optimize over a subset of the pixels. . 1 ! I C .. 75% of the pixels were excluded from the optimization. 5. ?' .. .Q e E 0. In conventional treatments.. . 2 4 6 Relative Dose -. % . From Figures 4. SHEPARD.4 i ! ! 1 0.. I '. 0.-. Thus. virtually all of the potential beamlets are used in the optimization. The field of radiotherapy is currently in a period of dramatic change. 4. typically the weighting parameters are changed and the amount of data involved is greatly increased. exactly match the prescription.6 " . : 8 10 (c) 15 beam angles (d) 71 beam angles Fig.. '.. it can be seen that a large number of beam angles are needed to successfully deliver conformal avoidance treatment plans. . \ ! ! . New treatment machines are providing the opportunity to treat cancer patients with thousands of independently weighted beamlets of radiation.7 and 4. MACKIE 0 .. 2 ! 8 I 10 4 6 Relative Dose (a) prescription (b) 7 beam angles . the majority of potential beamlets are removed from the optimization because they do not pass through the tumor. .. 0 . AND T. A computation point was placed at every other pixel in both the x and y directions.8 0. Thus. FERRIS.2 t. The optimization using 72 beam angles was performed over a grid of points. M. 1 ! \.. In conformal avoidance.8 A series of cumulative dose volume histograms corresponding to the conformal avoidance results shown i n Figure 4.. conformal avoidance simulations represent a particular challenge. From the perspective of optimization. .7.-. OLIVERA.. a large regional field is treated.. This is particularly true when the prescription calls for adjacent pixels to receive dramatically different dose levels.D. While the general form of the optimization problems is the same as those described above. The goal of this work has been to examine a number of mathematical models for . .8. .

B ~ R T F E LAND W. We believe that both model formulation and algorithmic improvements are necessary if problems of this size are t o be solved in an acceptable amount of time. J . The ideal optimization approach should provide the ability t o include partial volume constraints in an optimization model. MOHAN. 209-246. D. L. 123-135. Actual patient cases will involve as many as 100. AND Fitting of n o r m a l tissue tolerance data t o a n analytical function. CAROL. 1081-1089. 38 (1993)) pp. the beamlet data sets along with the code used t o compute them are available at the Web site http://www. KENDRICK. St. GOITEIN. E ~ I A ~ I I . eds. diother. . . B R A H ~ IO p t i m i z a t i o n of stationary and m o v i n g beam radiation therapy techniques. T. TARGOVNIK. pp. PAVORD. [3] T. p t i m i z a t i o n of beam orientations in radiation therapy: S o m e theoretical considerations. A. R. 129-140. Internat. BALOG.P resent and future developments in radiotherapy treatment u n i t s . Radiation Oncology Biol. and mixed integer approaches t o optimizing each patient's treatment plan. REFERENCES [I] J . pp.radiology. 5 (1995).madrad. Medical Physics Publishing. [6] A . The Scientific Press. T [8] L. BOYER. pp. RECKWERDT. 21 (1991). 146-155. wisc. J. L. Although these two approaches have proven successful for small-scale problems.. EDDY. 1995. South San Francisco. H. GRANT. 1997. BURMAN. Smith. ed. . D O [2] T.. BRAHME. [ll] M . . pp. . BUTLER. MACKIE. G [9] A . hree dim e n s i o n a l conformal treatment planning w i t h multileaf collimators. MAGERAS.OPTIMIZING THE DELIVERY OF RADIATION THERAPY 74 1 determining the optimal beam weights. S. It is our hope that the community of optimization experts will be able t o offer further insights that will improve our ability t o solve these difficult problems. CA.. either technique can become excessively time consuming in optimizing a real patient case. 373-375. BORTFELD. pp. KUTCHER. H B. Louis.. [5] A. G. Seminars in Radiation Oncology. X w i t h multileaf collimators.C . pp.Berlin. pp. BREWSTER. Starkshall. LEIBEL. the enormous amount of data involved in the optimization of conformal avoidance treatments necessitates substantial computational advancements. 33 (1995). nonlinear. The flexibility provided by NLP is particularly appealing. W D P.A N D P.r a y field compensation D. BROOKE. M.L . Springer-Verlag. R. S. R. J . FIGURA. Phys. SCHLEGEL. R. 723-730...T . ANGELOS.000 beamlets of radiation in a threedimensional setting. Woo. 291-304. E. Biol. Internat. J . ONUFREY. Salt Lake City. We have tested both a nonlinear and a mixed integer approach t o partial volume constraints. r e a t m e n t optimization: Using physical and radtobio~ogicalobjective functions. AMS: A User's Guide. J.G . BURMAN. 25 (1993). V.pp. [4] A .R. FUKS. B. 12 (1988). GROSSMAN. MEERAUS. 323-326. Med. [7] A. SELKAR. RaE. J .AND Z. D AND A . pp. Phys. MO. 28 (1994).edu. KAHLER. Radiation Oncology Biol. D. Radiation Oncology Biol. For readers interested in experimenting with these problems.. Internat. Radiation Oncology Biol. L. WALDRON. BOYER. Internat. Phys. B R A H ~ I Optamization of radiation therapy and t h e development of multileaf collimation. Our simulated treatments have helped t o further our understanding of the advantages and disadvantages of linear. R. Oncology. Stereotactic and Functional Neurosurgery. S. The use of nonlinear constraints in conjunction with a quadratic objective function should provide an oncologist with the flexibility t o produce an acceptable treatment plan for almost any disease site. J . 1988. . A number of critical challenges remain. c o m p u t a t i o n for a oneDose dimensional multileaf collimator. in Proceedings of the XI1 International Conference on the Use of Computers in Radiation Therapy. AND A . Similarly. 30-34. [ l o ] C. Phys. Phys. 66 (1996). Clinical implementation of an optimization approach requires a technique that is suitable for solving very large problems. Leavitt and G. AND Initial clanacal experaence wath t h e Peacock i n t e n s i t y m o d u l a t i o n of a 3-0 conformal radaation therapy s y s t e m . in T Radiation Therapy Physics.

Radiation Oncology Biol. pp.. 1997. 1981. Heidelberg. Springer Ser. C . L'Ied. [20] B . GUSTAFSSON. Med. [13] V. FERRIS. Biol. MACKIE. A L Scale Nonlinear Optimzzation (Release A ) . [36] P . versity of Wisconsin. .. Practical Optimization. London. S L Handbook. 128 (1978)) pp. Multiple local minima i n radiotherapy optimization problems with dose volume constraints. AND R . S. I.M A T L A B and G A M S : Interfacing Optimization and Visualization Software. Phys. D . Radiation Oncology Biol. R [22] M. 21 (1994). M. R . probability of radiation-induced complicaG. KALLIUIAN. 16 (1989). Internat. DEASY. FRAASS.R . A N D A.742 D. Phys. SVENSSON. [31] T. ALDRIDGE. B... DU. Phys. Med. N D H. ROM. Princeton University Press. 235-243. [14] A. Springer. p.. [15] R . BRAHIUIE. Phys. Med. A Simultaneous optimization of dynamic multileaf collimation and scanning patterns or compensation filters using a generalized pencil beam algorithm. R . Optimization of conformal treatment planning and quadratic dose objectives (abstract). CORMACK. R . GUSTAFSON. pp. ~ I U R R A AND M . C . 1645-1657. BURMAN. . A A tions for normal tissues wzth parallel architecture subject to non-uniform irradiation. RECKWERDT. R . 43 (1998). Phys. GOULD. A A. [34] A. HRISTOV. 91-106. G . MACKIE. B. J. 20 (1993)) pp. 1992. SHEPARD. 13 (1987)) pp. hIed. FERRIS. Med. NV. B. 332-334.AND T. H. SUTLIEF. CHVATAL. JACKSON. pp. 613-625. MARTINEZ. S. Math. D. Boyd and Fraser. KUTCHER N D C.. [27] P. Med. SYIulONS. 1998. D . uniform normal tissue irradiation: T h e effective volume method. J .A gradient method of optimizing external-beam radiotherapy treatment E plans. Registration of synthetic tomographic projection data sets using cross-correlation. Y. FITCHARD. NJ. [19] J . 1623-1630. 707-714. Radiation Oncology Biol. also available online from http://www. pp. YU. N D J . r\r. E . GRANT AND S. [30] T. pp. pp. Phys.filtered backprojection dose calculation method for inverse A E treatment planning. Med. Computer Sciences Department. FANG. Princeton. 9 (1999). comparison of three inverse treatment planning algorithms. C . [21] A.0. Medical Physics Publishing. [23] G . H . 208-223. YAN. Clinical and financial issues for intensity modulated radiation therapy delivery. CORIUIACK. TAYLOR. 623-630. Uni. LIND. pp. 22 (1995). Linear Programming and Extensions. [29] A. W. P AND T. BRAHME. [16] A. H. pp. Comput. MO. Software system for uw/ge tomotherapy prototype. H O L ~ I E S D T. [25] E. Verlag. A N C E L O T : A Fortran package for Large. GILL.. N D E. Academic Press. LEE. W O MA. 0 . Danvers. 1983. 1963. AND T. Phys. J . JACKSON. CHO. 62 (1992). MARKS. Radiation Oncology A Biol. GEISER. [24] hI. AND A. 249-262. S.cplex. . 21 (1994). Y. 39 (1994).E. pp. Louis. A. CONN. Med. pp. Salt Lake City. 25 (1998).W .. Radiology. pp. Radiation Oncology Biol. I. O H . pp. Phys. Phys. Tumor and normal tzssue responses to fractionated non-uniform dose delivery. M .C O N O P T : A G R G code for large sparse dynamic nonlinear optimization problems. L1Iedical Physics.com/.N D A. T 1921. 31 (1985). TOINT.A multzleaf collimator field prescription preparation system for conventional radzotherapy. J . pp. . [18] G. Seminars in Radiation Oncology. [26] A . St. 30 (1994). [35] A. DRUD. J. HUNG. H. PHILLIPS. J . Incline Village. 1911. Freeman. 1006. [17] A . A Optzmizatzon of intensity modulated beams with volume constraints using two methods: Cost function minimizatzon and projection onto convex sets. 22 (1995). WONG. S. Madison. 1997. Programming. 0 . Internat. Phys. 1141-1156. X. problem i n rotation therapy with z-rays. pp. L. New York. Internat.. S. 17. [32] ICI. [28] W. ILOG CPLEX Division. X. WRIGHT. FALLONE AND D. G. Phys. 1995. R . AN A Phys. WOO. M .. Phys. J . [33] C P L E X Version 6.S. WANG. KUTCHER.G . W. Internat. A Calculation of complication probability factors for non[37] G . MACKIE. .. Biol. 1921-1925. J . K . M . YORKE.~ I A C K I A . 153-191. MATTER. 303-313... MACKIE M [12] P . N D PH. continuous penalty function method for inverse treatA ment planning. Phys. WARREN. COOPER. Berlin. Lznear Programmzng. J.. 25 (1998). Mathematical Programming Technical Report 98-19. h e development of conformal radzation therapy. R. 99-107. . R. 435-443. OLIVERA. J . MOHAN. Internat. A problem i n rotatzon therapy with x-rays 2: Dose distributions with an azis of symmetry. in Proceedings of the XI1 International Conference on the Use of Computers in Radiation Therapy. DANTZIG.AGREN. HOLMES N D T. M . N. Math. 13 (1987).

J . Phys. \VAL.R . [61] A. R . New York. . 12 (1985). A K D T . NIELIIERKO. LEAVITT.20. Y. J . R J . Optim. CA. Med. R . [44] J. 1992. J . [53] S. BRE~VSTER. MACKIE. L/IORRILL.Clinically relevant optimazataon of 3-d conformal C. Med. G . J. LING. LYMAN histogram reduction algorithm. The Mathworks. T . RECKWERDT. MA. NAZARETH. Phys. .. KUTCHER. D. F [49] R . . Computer Solutzon of Linear Programs.. SWERDLOFF. R .. R . Radiation Oncology Biol. MD. convolution method of calculating dose E AN A from 15 mev x-rays. G L. J . 43-55.N D J ... MOREAND G . numerical study of the limited memory BFGS method and A the truncated-Newton method for large scale optzmzzatzon. Radiation Oncology Biol. T P A [46] T . through computer-controlled collimator motion and dose delivery. LtI. ROSEN. gramming. n the solution of large quadratic programming problems with O bound constraints. 1993. FUI<S. Phys. .. pp. 5 (1995). pp. A. 19 (1992).FUKS. pp. MURPHY. KIKSELLA. pp. AND Histogram reduction method for calculatang complicataon probabilities for three-dimensional treatment planning evaluation. KUTCHER. C . Philadelphia. FASS. ~ I O H A N . Seminars in Radiation Oncology. 6 (1995)) pp. 24 (1997). User's Guide.. G . Natick.D J . The American Cancer Society. J. C. Atlanta. TORALDO. C. Math. Phys..A N D Z. SIOHAN. [48] MATLAB. LANE. [59] J . NASHAND J. L'IoR~AND S. Med.AND Z. formal Therapy. PALIJ B. hIed. Phys. NATIOKAL IKSTITUTES HEALTH. M R L. MURTAGHK D M . [55] G . 433-436. pp. NOCEDAL. J. Int. LAM. Phys. Technical Report SOL 83. 179-188. C.. 49-55. [56] B .C. Phys. . J.. K A [54] S. 1709-1719. LEKHARD. [45] T . GOITEIN.. OF Radiation Therapy and You-A Guide to Self-Help During Treatment. BREIVSTER. [42] D.. 1995. J . Internat. 0. Imaging Systems and Technology. 17 (1990). 957-964. Frontiers Appl. . MARTIN. J. E . Phys. Dose-volume conszderatzons with linear pro. treatments. LLACER. S. 11. 87-92. 18 (1991).SI. Treatment plan optimization.. NIH Publication 97-2227.W. . LEIBEL. UK. M OHAK. [57] K .. 86-99. G . H O L ~ I E S. B. AKD EDS. BREWSTER. Radiation Oncology Biol. 1 (1991). Purdy and B. 137-146. R . Tomotherapy: A new concept for the delivery of dynamic conformal radiotherapy. Linear Programming. A.. RECKWERDT. J. I. comparason of mixed integer programmang and fast simulated A A annealing for optimized beam weights an radiation therapy. [43] J . J. N I E ~ I I E R K O . Ameracan Cancer Socaety Textbook of Clinical Oncology. in 3-D Radiation Treatment Planning and Con. A. pp. B. Inverse radiation treatment planning usang the dynamtcally penalazed lakelihood method. pp. HARRISON. MO. Optim. pp. Optimizataon of 3d radiataon therapy wath both physacal and biologacal end poants and constraints. [52] J . pp. 20 (1991)) pp. 45 (1989). J . BALDWIK. GA. G S. 1751-1764. ROSEN. pp. [60] A. Phys. [40] D. YANG. 1993. Stanford. h'ledical Physics Publishing. C. G B. WRIGHT. L. 1201-1210. MORRILL. HOLLIES. J . National Institutes of Health.. BURMAN. MOHAN. LA~VREKCE. n the limated memory method for large scale optimization. HUNT. Stanford University. J. pp. [58] S. [41] S. NOCEDAL.OPTIMIZING THE DELIVERY OF RADIATION THERAPY 743 [38] G. SIAhI J. Dynamic wedge field techniques M J. Emarni. 823-833. Int. 23 (1996). L. 20 (1993). [62] N. 99-108. W O K G AND I. J. L. eds. Improved dose distrzbutions for 3d conformal boost treatments i n carcanoma of the nasopharynx.Tomotherapy: Optimized plannang and delivery of radiation therapy. WOLBARST. T S P J. . LEIBEL. L. . . ~ I U R T Y .Very fast simulated annealing i n radiation therapy treatment plan optimization. H. MAGERAS. 1983. W . [50] R . KUTCHER. Phys. BURMAN. Oxford University Press. L. 503-528. ~ I A C K IJ. . Phys.0 User's Guade.. SAUKDERS. LANGER N D S. John Wiley. J. Louis. Med. 23 (1992). DEASY. . pp. Int. Phys. MORRILL. 3 (1991). A. E. O Programming. P . A MINOS 5. S. [47] T . SCRIIUIGER. R. 188-196. 1997. LING. Oxford. SIACKIE. C. LANGER. C . ield shaping for three-dimensional conformal radiation therapy and multileaf collimation. Radiation Oncology Biol. . W. 933-944. . I. Med. Med. 93-113. YANG. BATISTA. LAKE. Optimazation Software Guzde. LIU A N D J. K D I. 14. St. 31 (1995)) pp. [51] J. 358-372. 17 (1989)) pp. Internat. 1983. 11. SIAM. pp. [39] M . pp. 21 (1991). SIAM J . Math. ~ I O E L L EAND W .. D.. 1987. Bethesda. Internat. Radiation Oncology Biol. BURMAN. LEE. Optimazataon of radiation therapy iv: A dose volume AKD A.

P o i n t Methods.T r e a t m e n t plan optimization using linear A programming.. Phys. Biol. St. O p t i m u m parameters in a model of t u m o r control probability including interpatient heterogeneity. . FERRIS. D. Radiation Oncology Biol. J . H. pp. Phys. Investigation of t h e c o n t ~ o l u t i o nmeth. J . rznciples and Practice of Radiotherapy. SANDERS. 1997. OLIVERA. LANE. P . St. J . pp. MORRILL. A. Medical Physics Publishing. Medical Physics Publishing. ANGELOS. 100-113. hI. . [76] S. PEREZ AND L. R . RECKWERDT. AND Megavoltage C T o n a t o m o t h e r a p y s y s t e m . D B.. FITCHARD. R. . D. P Raven. PAPANIKOLAOU. 1997. . BELLI. Advanced Medical Publishing. Med. 1968. Seminars in Radiation Oncology. 5 (1995). RECKWERDT.. RECKWERDT. BRADY. [70] D. [75] S. GRANT. Med. Phys. 1327-1336. 0 . P [77] J . H. GEHRING. Phys. t o appear. simple model for examining L. Phys. SCHLOESSER. N. N D E. .R .J . 1998. invesT P J AN An tigation of t o m o t h e r a p y beam delivery. 420-422. Y. MACKIE. 213-214. A. Med. tice of Intensity Modulated Radiation Therapy. 24 (1997). UT.C. [65] N. J . WI. Philadelphia.. 0 A A radiotherapy optimization. T h e Physics of C o n f o r m a l Radiotherapy-Advances in Technology. T [67] P . R. SHEPARD. [74] S. WRIGHT. A Does t h e "peacock" have anything t o d o w i t h radiotherapy?. R. R . 1895-1914. 327773294. AND T. 29 (1994). pp. MEGER-WELLS. SAUER. ZACHMAN. B. MO. . OLIVERA. [73] S.A N D T . W. J . The Theory and Prac. 141-152.744 D. MACKIE. Phys. M. W o o . ROSEN. W.Inverse Planning for I M R T : T h e Role of Simulated Annealing. Phys. pp. Institute of Physics Publishing. Biol. PA. SHEPARD. 11 (1994). Salt Lake City. rimal-Dual I n t e r i o r . 20 (1993). Biol. IC1cGrau:-Hill. Starkshall. THOMADSEN. MO. R .od for polyenergetic spectra. 1997. T.. . LtIed.. [69] K . Madison. Internat. M. MACKIE [63] G. RUCHALA. [71] L. 425-436. MACKIE. Louis. P K J E.. pp. R . MACKIE.. BALOG. Salt Lake City. MACKIE. N D J . S. UT. N D T . ORCHARD-HAYS. Louis. 1996. WEBB. MCNUTT. A . LtI. [66] C. Philadelphia. AND T. Advanced Linear Programming C o m p u t i n g Techniques. RECKWERDT. T. M . WEBB. eds. New York. SIAM.. Leavitt and G. DEASY.. E. MACKIE. RUCHALA. 1997. 3rd ed. a x i m u m likelih~ood as a c o m m o n computational frame. Med. OLIVERA. G E. [72] S. YANG. J . in XI1 International Conference on the Use of Computers in Radiation Therapy. AND P . . in Proceedings of the XI1 International Conference on the Use of Computers in Radiation Therapy. Philadelphia. R . G. D. M work in tomotherapy. Lippincott. SHEPARD. [68] I. VERHEY. pp. [64] W . Immobilizing and positioning patients for radiotherapy. 43 (1998)) pp. Three-dimensional i n v e r s e treatment o p t i m i z a t i o n for tomotherapy.. hIed. WEBB. BUTLER. 18 (1990). pp.

- test
- ECG MONITORING OF A CARDIAC PATIENT USING EMBEDDED SYSTEM
- MRI Trends 1
- aigh
- Tata Evidence Based Medicine - Modern Radiation Oncology Practice..pdf
- h Stasis
- Report 3 Ack
- THR Content Specification
- Review Ductus Arterious
- The Use Picco in Unit Care
- Blood Analysis Without Pain for Neonates
- IL Ch08 Transport in Humans
- Collection (1)
- Use of Extraoral Periapical Radiography in Indian Population
- CT Ultrasound Registration
- Nephrectomy and Spill
- +Multi-Detector Row Computed Tomography- Imaging the Coronary Arteries
- Brachytherapy GYN Applicator Catalog
- Glaucoma incidence in an unselected cohort of diabetic pts.pdf
- Zap at A
- dbs-kin7428-tchmanl
- SLE 5000
- 2371
- Exogen 4000+ Instructions for Use
- Surgery 542 Trauma
- Acta Radiol 2011 Al Saeed 570 4
- Update on Device Technologies for Monitoring HF
- Sigmoidal Equation for Lung and Chest Wall Volume Pressure Curves
- Gastric Bypass Outline
- Mobile Phone Radiation Health Risk Controversy

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd