Controversies_in_Medical_Physics_a_Compe.pdf

CHAPTER 1 General Radiation Therapy 1.1. Monte Carlo techniques should replace analytical methods for estimating dose distributionsin radiotherapy treatment planning Radhe Mohan and John Antolak Reproduced from Medical Physics, Vol. 28, No. 2, pp. 123–126, February 2001 ( 001&idtype=cvips&gifs=Yes) OVERVIEW Analytical models have traditionally been used to estimate dose distributions for treatment planning in radiation therapy. Recently, some physicists have suggested that Monte Carlo techniques yield more accurate computations of dose distributions, and a few vendors of treatment planning systems have incorporated Monte Carlo methods into their software. Other physicists argue that, for a number of reasons, analytical methods should be preserved. This controversy is the topic of this Point/Counterpoint article. Thanks are extended to Paul Nizin, Ph.D. of Baylor College of Medicine forsuggesting the topic. Arguing for the Proposition is Radhe Mohan, Ph.D. Dr. Mohan received his Ph.D. from Duke University and is currently Professor and Director of Radiation Physics at the Medical College of Virginia (MCV) Hospitals, Virginia Commonwealth University. Dr. Mohan has been actively engaged in research and clinical implementation of advanced dose calculation methods, 3D treatment planning, Monte Carlo techniques and IMRT for 25 years, first at Memorial SloanKettering Cancer Center and now at MCV. He has published and lectured widely on these topics at national and international meetings and symposia. Arguing against the proposition is John Antolak, Ph.D. Dr. Antolak received his Ph.D. in Medical Physicsfrom the University of Alberta (Canada) in 1992. He then joined the Department of Radiation Physics at The University of Texas M. D. Anderson Cancer, where he is currently an Assistant Professor. He is certified by the American Board of Radiology and licensed to practice Medical Physics in Texas. He is active in the education of graduate students, dosimetrists, and other physicists, and his research interests center around the use of electron beams for conformal radiotherapy. In his spare time, he enjoys playing ice hockey and coaching his son's ice hockey team. FOR THE PROPOSITION: Radhe Mohan, Ph.D. Opening Statement 2 Monte Carlo techniques produce more accurate estimates of dose than other computational methods currently used for planning radiation treatments. Were it not for limitations of computer speed, Monte Carlo methods probably would have been used all along. With the spectacular increase in computer speed and the development of clever algorithms and variance reduction schemes, Monte Carlo methods are now practical for clinical use. The time required to compute a typical treatment plan has shrunk to a few minutes on computers costing less than $50,000. A few centers have started using Monte Carlo techniques for clinical purposes, and releases of commercial products are imminent. As with any new product, an "adjustment period" will be needed during which we learn how to apply this powerful tool. Some find the "statistical jitter" in Monte Carlo results troubling. This issue is being addressed by several investigators. The additional cost of hardware and software may be another obstacle, but is likely to be resolved as computers become cheaper and more powerful. Another issue is whether improvements in accuracy are clinically significant and worth the additional cost. It is difficult to answer the first question unequivocally, because randomized trials in which half the patients are treated with less accurate methods are not feasible. Arguments in favor of using Monte Carlo methods include: (1) Elimination of the need to continually reinvent approximate dose computation models and to tweak them to meet every new situation, as well as the need for trial and error approaches to obtain acceptable matches with measured data. The medical physics community has been engaged in such exercisesfor 50 years. It is time to stop. (2) Broad applicability and accuracy of the same Monte Carlo model for all anatomic geometries and treatment modalities (photons, electrons, brachytherapy). With analytical methods, there is a separate model for each modality and a unique set of approximations and assumptions is required for each type of field shaping device. (3) Dramatic reduction in the time, effort and data required for commissioning and validating the dose computation part of treatment planning systems. (4) Improved consistency of inter-institutional results, and greater quality of dose response data because of improved dose accuracy. (5) Accurate estimation of quantities difficult or impossible to measure, such as dose distributions in regions of disequilibrium. Until recently, the major reason for considering Monte Carlo methods was the inaccuracy of semi-empirical models for internal inhomogeneities and surface irregularities. Now an equally important justification is the ability of Monte Carlo techniques to provide accurate corrections for transmission through, scattering from, and beam hardening by field shaping devices. Monte Carlo techniques are also able to account correctly for radiation scattered upstream from fieldshaping devices. These effects are quite significant for small fields encountered in intensitymodulated radiotherapy. The transition to Monte Carlo methods will have to be gradual. Even though a few minutes of time to compute a plan may seem insignificant, computer-aided optimization of treatment plans may require many iterations of dose computations. In these situations, hybrid techniques will be needed that use fast but less accurate conventional models for most optimization iterations and Monte Carlo techniques for the remainder. Since Monte Carlo techniques are now affordable and practical, there is no reason not to use them. It is not necessary to conduct clinical trials to once again prove the clinical significance of 3 improved dose accuracy. Monte Carlo methods should be deployed in radiation therapy with deliberate speed. For some applications, such as, IMRT optimization, it may be necessary to continue to use high-speed conventional methods in conjunction with Monte Carlo techniques at least for now. Rebuttal Dr. Antolak has raised several issues, some of which were addressed in my Opening Statement. With faster computers and clever schemes to reduce variance, the stochastic nature of the Monte Carlo approach is no longer an impediment. Statistical uncertainty of 1%–2% is achievable on grid sizes of 2–3 mm in MC dose distribution calculations, requiring just a few minutes on easily affordable multiprocessor systems. While statistical noise may be unsightly, its effect on the evaluation of dose-volume and dose-response parameters of plans is insignificant. In addition, techniquesto smooth out noise are being implemented. Analytic models introduce systematic errors in dose. They simply cannot achieve the accuracy of MC techniques. While it is true that analytic models consistently produce precise results for the same input data, these results are consistently inaccurate. Dr. Antolak is concerned that approximationsto speed up Monte Carlo computations may affect the accuracy of results. But Monte Carlo developers and users should always ensure that approximations have no significant impact on accuracy. Nothing else should be necessary. Responses to other such concerns raised by Dr. Antolak are: (1) Considering the uncertainties in dose-response information and other sources of data in the radiotherapy chain, our ability to define "how much noise in the dose distributions is acceptable" is similar to our ability (or lack thereof) to determine the level of dose inaccuracy that may be acceptable. (2) Dose to a point is not a meaningful quantity when Monte Carlo techniques are used. Beam weighting and dose prescription should be specified in terms of dose to fractional volumes (e.g., 98% of the tumor volume). (3) Statistical noise should have practically no effect on inverse treatment planning because the intensity along a ray is affected by the average of dose values over a large number of voxels lying along the ray and not by the dose in any one voxel. (4) Commissioning of Monte Carlo algorithms will be the responsibility of the same physicists and/or commercial vendors who commission conventional methods. I believe strongly that concernsraised by Dr. Antolak and others are being resolved and that we are now ready to introduce Monte Carlo techniques into clinical use. AGAINST THE PROPOSITION: John Antolak, Ph.D. Opening Statement We have a professionalresponsibility to ensure that patient treatments are accurately delivered, and the accuracy of treatment planning dose computation is one aspect of this. There are data to support the conclusion "that Monte Carlo techniques yield more accurate computations of dose distributions," provided that the Monte Carlo technique is fully applied. However, in light of other factors detailed below, Monte Carlo methods should not replace analytical methods for estimating dose distributions. 4 Before arguing against the proposition, it is necessary for me to clarify what I believe the basic difference is between an analytical method and a Monte Carlo method. It boils down to the difference between deterministic and stochastic. The Monte Carlo method is stochastic, i.e., independent calculations of the same problem will give different answers. The analytical method is deterministic, i.e., independent calculations of the same problem will give the same answer, at least to within the limits of numerical round-off and truncation errors. In my opinion and for the purpose of this discussion, any nonstochastic method is considered to be an analytical method. The accuracy of an algorithm (or method) describes how close it comes to the true answer. Clinical physicists have to worry about the accuracy of both analytical and Monte Carlo methods. The full Monte Carlo method (e.g., EGS4) is considered by many to be the gold standard for accurate dose calculations. The precision of an algorithm is a measure of the repeatability of the answer. Analytical methods have essentially absolute precision. However, the precision of the MonteCarlo method, as measured by the standard error, is proportional to the inverse of the volume of the dose voxels, and to the inverse square root of the computational resources allocated to the problem. For example, reducing the standard error by a factor of two requires four times as muchCPU-time. Variance reduction techniques can be used to reduce the computational resources required to obtain a given precision. However, the time (or resources) required for full Monte Carlo simulations of patient dose distributions is currently too great for clinical use. By necessity, current Monte Carlo treatment planning algorithms (those being touted for clinical use) introduce approximations that greatly speed up the calculations, but the accuracy of the results may be affected. At the same time, significant improvements are also being made to the accuracy of analytical algorithms. Also, for the clinical physicist, commissioning analytical algorithms is relatively straightforward, noise is not a problem, and the accuracy can be easily documented. From the perspective of the clinical physicist, many questions about the use of Monte Carlo algorithms have not yet been answered. How much noise in the dose distributions is acceptable? In the presence of noise, how should beam weighting (e.g., isocentric weighting) be done? What effect does noise have on inverse treatment planning? Who will take responsibility for commissioning the algorithm, and how accurate are the results of the commissioning? How long will the calculation take relative to faster analytical calculations? How will the calculation time affect treatment-planning throughput, particularly when using optimization methods? Is the spatial resolution sufficient for clinical use? Most of the time, Monte Carlo treatment planning calculation times are quoted for relatively coarse (e.g., 5 mm) spatial resolution. Just reducing the resolution from 5 mm to 3 mm requires approximately five times as much CPU-time. These are just some of the issues that need to be resolved and well-documented before Monte Carlo methods can replace analytical methods for treatment planning. Monte Carlo methods may be used as an independent verification of the dose delivery, or to document (rather than plan) the dose delivery. However, until the questions above are successfully answered, Monte Carlo methods should not replace analytical methods for estimating radiotherapy dose distributions. Rebuttal 5 We agree that greater accuracy in dose computation is desirable, Monte Carlo methods can produce more accurate dose estimates, and "Monte Carlo methods should be deployed in radiation therapy with deliberate speed." However, these points are not the proposition we are addressing. A potential patient recently inquired about the status of Monte Carlo planning at our institution. From what he had read, he believed that Monte Carlo treatment planning is a "silver bullet." Dr. Mohan says it is time to stop reinventing. I believe that implementation of "clever algorithms and variance reduction schemes" is reinventing Monte Carlo treatment planning methods. Further, trial and error will not stop with Monte Carlo. With complete information about source and machine geometry, Monte Carlo calculations can be highly accurate. However, Monte Carlo algorithms usually start with a source model that requires trial and error adjustments to match measured data. Whereas reinventing analytical methods usually improves accuracy, reinventing Monte Carlo methods may decrease accuracy. In both cases, there is a tradeoff between accuracy and speed, which is often seen if the Monte Carlo approach averages the dose over large voxels. How is the accuracy of a particular implementation judged and to what should it be compared? The "spectacular increase in computer speed and the development of clever algorithms" noted by Dr. Mohan permits significant improvements in analytical models, potentially leading to a model for coupled photon-electron transport under more general conditions. Future reductions in commissioning and validation efforts will only come from manufacturers' standardization of treatment machines and improved quality in their construction. Modified and new machines will still require extensive commissioning and validation for both Monte Carlo and analytical methods. Considerable research remains to be done to identify the minimum data set sufficient to validate input data that characterizes a treatment machine. Dr. Mohan's last two points are really arguments for greater dose accuracy and apply to treatment planning systems in general, not just Monte Carlo methods. Dr. Mohan cites the significance of Monte Carlo methods applied to field shaping and intensity modulation devices. These applications can be very complex, but are usually not modeled explicitly. For example, modeling all of the field segments for a complex DMLC fluence pattern is impractical under normal circumstances. Using an approximate approach affects the overall accuracy of the dose calculation, as it would for an analytical method. Monte Carlo methods are an invaluable tool for improving analytical models to a point where their dose uncertainty is insignificant compared with other uncertainties radiation therapy—such as setup, internal organ motion, target delineation, and biological response. As stated earlier, "Monte Carlo methods may be used as an independent verification of dose delivery, or to document (rather than plan) dose delivery," but should not replace analytical methods for estimating dose distributions in radiotherapy treatment plannin