Journal of Physics: Conference Series

PAPER • OPEN ACCESS Related content

- Computer-aided design optimization with
3D dose computation algorithms the use of a fast dose model for linear-
accelerator-based stereotactic
radiosurgery
To cite this article: T Knöös 2017 J. Phys.: Conf. Ser. 847 012037 Tae S Suh, Frank J Bova, Sei C Yoon et
al.

- Influence of output measurement time

interval and tolerance on dose deviation
Mika Kapanen, Mikko Tenhunen, Ritva
View the article online for updates and enhancements. Parkkinen et al.

- Estimation of neutron-equivalent dose in

organs of patients undergoing
radiotherapy by the use of a novel online
digital detector
F Sánchez-Doblado, C Domingo, F
Gómez et al.

This content was downloaded from IP address 111.68.99.208 on 02/11/2017 at 03:14

9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

3D dose computation algorithms

T Knöös1,2
1
Department of Haematology, Oncology and Radiation Physics, Skåne University
Hospital, SE-205 02 Malmö, Sweden
2
Department of Medical Radiation Physics, Clinical Sciences, Lund University,
SE-221 85 Lund, Sweden

Email: tommy.knoos@mod.lu.se

Abstract. The calculation of absorbed dose within patients during external photon beam
radiotherapy is reviewed. This includes the modelling of the radiation source i.e. in most cases a
linear accelerator (beam modelling) and examples of dose calculation algorithms applied within
the patient i.e. the dose engine. For the first part - the beam modelling, the different sources in
the treatment head as target, filters and collimators etc are discussed as well as their importance
for the photon and electron fluence reaching the patient. The consequences of removing the
flattening filter, which several vendors now have made commercially available, is also shown.
The pros and cons regarding different dose engines ability to consider density changes within
the patient will is covered (type a and b models). Engines covered are, for example, pencil-beam
models, collapsed cone superposition/-convolution models and combinations of these, as well as
a glimpse on Monte Carlo methods for radiotherapy. The different models’ ability to calculate
dose to medium (tissue) and or water is. Finally, the role of commissioning data especially
measurements in today’s model based dose calculation is presented.

1. Introduction
Clinically used dose planning systems have for many years used calculation algorithms for X-rays and
γ-beams, which make use of empirically, determined inhomogeneity corrections. These corrections have
been applied to either measured dose distributions using e.g. film or simple analytical models describing
the distribution in homogeneous water. Modelling of the dose distribution has not relied on basic photon
and electron interaction (basic principles). The methods in use today are often model based algorithms
where one separates the modelling of the radiation source from the in-patient dose calculation, the so-
called dose engine [1].

2. Historical algorithms
A short description of older algorithms seems to be in place and we divide then depending on their
ability to model scattered radiation. These models can be summarised as correction models applied to
the dose distribution in water.

2.1. Algorithms without Scatter Modelling

In this group of algorithms, e.g. effective attenuation, isodose shift, tissue air ratios (TAR), effective
source skin distance (SSD) [2] the only influence of the patient geometry is thru a scaling of the
radiological depth for the calculation of absorbed dose. More or less sophisticated versions exist -
approximations of homogeneous media, regional corrections or pixel by pixel corrections. These

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

methods only account for inhomogeneities with respect to their density along the fan line. A better
method was introduced with the algorithm by Batho where the position of the inhomogeneity also is
considered [3]
These methods were all, in principle, developed prior to the use of CT. The resulting dose
distributions from these models can in some clinical cases differ up to 20 % from measurements,
especially in low density media irradiated with narrow (5 x 5 cm2) high energy photons beams [4].
Deviations of the same magnitude between measurements and calculations were found for a tangential
beam geometry using the effective attenuation method. Of these methods, the Batho method estimates
the absorbed dose with the highest accuracy. Lulu and Bjärngard have extended the Batho algorithm to
account also for the lateral extension of inhomogeneities [5, 6].

2.2. Algorithms with Scatter Modelling

The next group of methods all use the radiological depth to determine the primary dose component but
also accounts for scatter produced in the irradiated volume.
The Equivalent Tissue Air Ratio, ETAR algorithm relies on a calculation of an effective
homogeneous density which is assigned to each point using weighting factors applied to the surrounding
points (voxels) [7]. Geometric dimensions (e g depth, field size) are scaled from unit density to the
effective density using O'Connor's theorem [8, 9]. The weighting factors used are calculated using
inelastic scattering cross sections for water, i.e. basic principles are applied. The agreement with
measurements are better for this method than for the simple algorithms. An accuracy of ± 3 % can be
expected in most cases [7], however, in some situations, the deviation for the ETAR method can
increase, for example, tangential beams, where deviations up to 5 % can be found [10].
An extension of the ETAR method using TAR's calculated by Monte Carlo have been proposed
[10]. The new TAR's are more accurate especially for small fields where they fall to zero for zero field
size. Therefore, the absence of lateral electronic equilibrium for narrow beams in low density media is
modelled better. The final equation used is a superposition integral similar to the methods discussed by
Mackie et al [11] and Mohan et al [12].
In general, none of the methods discussed deals with the lack of electronic equilibrium for narrow
beams of high energy X-rays in low density media. All these models assume that the energy transferred
to electrons is absorbed locally i.e. the collision kerma is equal to the absorbed dose.

3. Modern dose calculation algorithms

3.1. Multi source models

The limited accuracy of the methods described above has initiated much work on methods based on
physical principles. The approach used mainly today is fluence or energy fluence multi source models.
The latter is describing the radiation source, usually the treatment head of the accelerator down to
different detailed levels. For example, the target generating x-rays from the impinging electron which is
considered the primary radiation beam, scattered radiation from primary collimator, beam flattening
filter (if present), monitor chamber, multi leaf collimator and conventional jaw system. A common
simplification is to reduce all scatter sources to one single usually placed at the position of the flattening
filter. This position is chosen since the flattening filter accounts for the majority of the scattered radiation
from the treatment head.
Once the model of the radiation source is set up, one calculates the output from it, as an energy
fluence distribution. In the case of using Monte Carlo this can be a phase space of the energy fluence
i.e. a description of type of particles, their position, direction, and energy. The energy fluence
distribution is modulated by field shaping parts of the treatment head. The sources lateral distribution
and relative magnitude is usually adjusted based on measurements of conventional measurements e.g.
lateral profiles, depth doses, output factors in air and water.

2
9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

Figure 1. A schematic describing the separation of the

modelling into a multi-source model for the treatment head and
an in-patient dose calcualtion. The plane below the jaw system
represents the phase space of out-coming particles. The patient
is represented with a single transversal cross-section and
isodose lines are added for illustration.

3.2. In-patient or dose engine

Dose calculation is based on the previous energy fluence impinging on and into the patient. The patient
is represented by a 3D voxel matrix where each voxel is representing the density (relative physical or
electron density depending on system). For some models, also a medium may be given for each voxel.

3.2.1. Point kernels

The energy fluence is ray traced thru the patient matrix and attenuated considering the density of each
voxel and when appropriate also the medium of the voxel. For each voxel, the TERMA (Total Energy
Released in Mass is calculated as the product of the energy fluence Ψ and the mass energy absorption
coefficient µ/ρ. The following equation gives the TERMA T differential in energy E at an arbitrary point
in the 3D room given by the vector r.
𝜇𝜇𝐸𝐸
𝑇𝑇𝐸𝐸 (𝒓𝒓) = ∙ Ψ𝐸𝐸 (𝒓𝒓)
𝜌𝜌
From the TERMA distribution one gets the absorbed dose by a convolution with a point kernel
describing the transport of energy by primary electrons release (collision kerma) and scattered photons.
(The differential in energy has been excluded for clarity, the integral has to be performed over all
energies).

𝐷𝐷(𝒓𝒓) = � 𝑇𝑇(𝒓𝒓′ ) ∙ 𝑃𝑃(𝒓𝒓 − 𝒓𝒓′ )𝑑𝑑𝒓𝒓

The differential in energy has been excluded for clarity, the integral has to be performed over all
energies. It is, however, quite common to perform the integral over energy before the convolution, thus
only one TERMA matrix and one single point kernel is required.

3
9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

Figure 2. A matrix of TERMA (dotted lines) is

convolved with a point kernel, shown as isolevels
(solid lines and broken lines respectively) representing
the pattern of absorbed dose around an interaction
point. The absorbed dose to the shaded voxel (dose
point) is determined as the TERMA at the interaction
point times the level of the kernel at the dose point.
This is repeated for all interaction sites (only two is
shown for clarity) to get the total absorbed dose to a
single voxel.

This time-consuming integral can be solved using Fourier transforms where the convolution
integral is replaced by the inverse Fourier transform of the product between the transforms of T and P.
This approach is fundamental in the work by Boyer who used Fast Fourier Transform (FFT) to solve the
convolution of dose distribution kernels with photon fluence distributions to give the final dose in 3D
[13, 14]. However, if the kernel function P varies with position due to changes in density, an analytical
superposition integral has to be solved instead.
The point dose kernels used, which includes electrons released in the first interaction as well as
single and multiple scattered photons, are generally calculated using Monte Carlo simulations but
analytical methods are also applicable.

3.2.2. Convolution in inhomogeneous medium

Inhomogeneous media can be included in the convolution integral either using a large number of dose
distribution kernels, one per density, or a single kernel combined with the following scaling theorems:
"In a medium of a given composition exposed to a uniform flux of primary radiation (such as X-rays
or neutrons) the flux of secondary radiation is also uniform and independent on the density of the
medium as well as of the density variations from point to point.” [15]
"In a given irradiation system the ratio of scattered photons to primary photons remains unchanged
when the density of the irradiated material is changed if all the linear dimensions of the system are
altered in inverse proportions to the change in density.” [8, 9]
Density in this context should be interpreted as interaction sites per unit volume, e g electron
density for incoherent scattering of photons. It is also assumed that the amount of secondary radiation
produced at a point is proportional to the local number of interaction sites.
Mackie et al used superposition kernels calculated for several densities with interpolation and
scaling between them [11]. A similar approach using kernels for only a single density combined with
the scaling theorems has been applied [12, 16-18].They all use ray tracing of scatter between interaction
and absorption point and employ the scaling theorem for inhomogeneities. Transportation of energy by
electrons is included in the kernels, therefore, electronic disequilibrium will be modelled.

3.2.3. Pencil beam kernels

A method to increase the calculation speed to make it clinical feasible was the introduction of pencil
beam kernels where the dose calculation problem has decreased by going from 3D to 2D convolution.
By integrating the point kernel along the depth direction, we get a 2D description of the dose distribution.

4
9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

1.0
0.9
0.8
Relative fluence

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0 10 20 30 40 50
Depth (cm)
Figure 3. Illustration of the principle of pencil beams showing creation of pencil beams by
convolving/integrating TERMA along a ray with the point kernel describing the energy transport and
absorption. Left panel - TERMA along the central ray of a very narrow photon beam, middle panel - point
kernel and right panel - the resulting pencil beam.

Pencil beams have been created in various ways for example from measurements by radial
differentiation of relative dose on central axis from broad beam dosimetric quantities [19]. Another
method was the differentiation of radial beam data [20]. Monte Carlo methods have also been used to
calculate pencil beam kernels in water [16, 21, 22]. Usually the pencil beams are parametrized and the
actual values of these are found by a fitting process to measured data [22].
Applying the pencil beam convolution is described in principle by the following equation:

𝐷𝐷(𝑥𝑥, 𝑦𝑦, 𝑧𝑧) = � Ψ(𝑥𝑥 ′ , 𝑦𝑦′) ∙ 𝑃𝑃(𝑥𝑥 − 𝑥𝑥 ′ , 𝑦𝑦 − 𝑦𝑦 ′ , 𝑧𝑧)𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑

Here the energy fluence distribution Ψ at a certain specified plane is convolved with the pencil
beam kernel P. In heterogeneous media, the pencil beam kernel is scaled along the propagation direction
by replacing the z with the radiological depth zradiol considering the density of the voxels along the ray.
No scaling is performed, in the initial implementations, perpendicular to the propagation direction. Thus,
the pencil beam convolution is a so-called type a algorithm [23]. There is, however, today one pencil
beam model where a lateral scaling of the kernel has been added [24]. The commercial implementation
is the analytical anisotropic algorithm (AAA) [25, 26]. The performance of pencil beams algorithms
have been reported in several reports [27-32].

Figure 4. Illustration of the 2D energy fluence distribution (left) convolved with the pencil beam (middle) to
get the dose distribution represented by a tranvese plane (right).

5
9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

3.2.4. Collapsed cone convolution

Another approximation to increase speed (which maybe is not a problem today) as well as managing
inhomogeneities is to discretize the point kernel. The kernel is discretized in about 100 directions and
by letting each direction representing a cone where all energy is collapsed to the center of the cone [33,
34].

Figure 5. The point kernel (illustrated by the isolevels) is discretised

along the arrows which each represent a cone where energy is
transported. All energy in the cone is collapsed to the central ray.

When collapsing the energy in the cone into the central ray the inverse square dependence is
removed. The distribution of cones is not isotropic, instead a higher density is used in the forward
direction because the majority of energy in high energy beams is transported in this direction. A few
cones take care of the backscatter. Density of cones should be high enough where significant amount of
energy is deposited such that each voxel is passed by one or more collapsed cones.
The convolution integral is commutative thus changing the order of the functions does not change
the result
𝐷𝐷(𝒓𝒓) = ∫ 𝑇𝑇(𝒓𝒓′ ) ∙ 𝑃𝑃(𝒓𝒓 − 𝒓𝒓′ )𝑑𝑑𝒓𝒓 is equal to 𝐷𝐷(𝒓𝒓) = ∫ 𝑃𝑃(𝒓𝒓′ ) ∙ 𝑇𝑇(𝒓𝒓 − 𝒓𝒓′ )𝑑𝑑𝒓𝒓

This property can be interpreted as shown in the following figure:

Figure 6. Illustration of the commutative property of the convolution integral. The left panel shows the "dose
deposition view" and the right panel the "interaction view" [4].

6
9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

For the dose deposition view, energy or TERMA from the surrounding voxels is summed up to
the deposition voxel. When all contributions are summed up, the total dose in the voxel is determined.
In the interaction view, the TERMA in a voxel is spread out to the surrounding voxels and all
contributions from interaction points have to be covered. In principle, the first approach can be used to
calculate the absorbed dose in a single point/voxel only. The implementations based on this approach
utilize the possibility to have a dose grid with different spacing to speed up calculations. For example,
in areas with small gradients in TERMA or patient density a coarse grid can be used and in high gradients
a finer one.
Another approach used to efficiently perform the convolution is to align the point kernels such
that the cones in the same direction overlaps. This makes it possible to ray trace through the TERMA
matrix along each cone direction and recursively pick up energy from each traversed voxel, transport,
attenuate and then deposit energy along the axis of the cone [33].

3.3. Linear Boltzmann transport equation

A more direct way of performing in-patient dose calculations is to start with the linear transport equation
by Boltzmann (LBTE). For many years one have used Monte Carlo (MC) techniques to statistically
solve the problem [35-38]. MC methods use stochastic distributions describing the physics to simulate
the transport of individual particles (photons and charge particles) through the density and medium
matrix describing the patient geometry. The limitation, at least up today, is the large number of particles
is required to get a sufficient high uncertainty of the deposited dose in each voxel of interest. MC can
be seen as a microscopic solution of the LBTE.
Recently analytical solutions have been applied to LBTE for radiotherapy planning problems in
a macroscopic approach i.e. instead of individual particles we create, follow, and attenuate energy
fluencies in a matrix describing the patient. An exact analytical solution is not possible, mainly due to
its complexity, thus a numerical have to be used, i.e. discrete ordinate method or grid based solvers [39-
42]. Evaluation of a commercial implementation have been presented by e.g. Fogliata et al [43, 44] and
Hoffman et al [45].
In short, the grid based method can be described as; the properties of the radiation source and the
absorbing medium (patient) are described, the multi-source model (e.g. by a phase space) is used to get
the energy fluence to the patient and then ray traced through the irradiated volume (observe this is very
similar to the point kernels methods discussed above). In the next step the discrete ordinates is applied
to solve the LBT equation giving photon energy fluence and the electron sources in each voxel. Finally,
the electrons are transported to obtain electron fluence distribution in each voxel and by applying mass
stopping power the absorbed dose is determined. The energy of photons and electrons are discretized,
25 and 49 levels, retrospectively. Transport directions is discretized based on the energy involved
between 32 and 512 directions (These data are from the Acuros implementation [46]).
Interesting to comment is that both solutions to the LBTE show inherent errors, for MC we have
stochastic uncertainties. For discrete ordinate solvers, the concept of discretization may produce
uncertainties. As always this can be improved by prolonging the computation by increasing the number
of particles in the MC case and for the solver using a finer discretization.

4. Dose to medium or dose to water

For many years, there have been a discussion on how to report absorbed dose during radiotherapy
procedures [47]. No consensus really exists if the dose should be reported as absorbed in water or in the
actual medium. With new algorithms as MC and grid based LBTE solvers getting more common this
problem is growing. Both these approaches can report either.
The issue can be divided into two parts, a) if the energy/TERMA during the ray trace considers
the physical properties of the transversed voxels or all media is considered to be water, and b) the other
step is how the deposition of energy is accomplished. Introducing the formalism
Dtransport, deposit where transport and deposit are the medium used respectively. They can either take the
value water (w) or medium (m).

7
9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

Considering the kernel approaches, pencil beam and point kernels, are all determined in water,
analytical, MC or from experiments. Looking in detail to pencil beam algorithms, scaling according to
the transversed density (radiological depth scaling) assumes all media is water and no corrections are
performed during dose deposition. Thus pencil beam models transports and deposit dose in water, Dw,w.
For the point kernel models available, no consistent handling exists. In principle, the TERMA or
equivalent can be ray-traced considering the medias physical properties or not. The kernels are
determined in water but when dose deposition occurs a correction using mass stopping power ratio
between medium and water may be applied. No single conclusion on what is reported can be given,
however, we probably have one of the following situations: Dw,w, Dm,m or Dm,w.
MC models have the possibility to transport in water or medium and the same is valid for dose
deposition. Most common is to transport in medium and deposit in medium. If the user want dose to
water, a mass stopping power ratio between water and medium is applied to each voxel [48]. Thus we
have Dm,m or Dm,w. When applying the mass stopping power ratio one usually have to use a macroscopic
value for the energy deposit in the voxel since the electron energy fluence differentiated in energy in the
voxel is not known. Alternatively, this could be done microscopically if done for each single energy
deposition.
The discrete ordinate solvers for the LBTE determines the electron energy fluence differentiated
in energy for each voxel and either the mass stopping power for the medium or for water is used to
determine the absorbed dose. All transport is accomplished in the medium thus we have Dm,m or Dm,w.
The choice is user selectable in the only existing commercial implementation [49, 50].

5. Summary
Today we have very accurate tools available for estimating the absorbed dose within the patient during
radiotherapy [51-57]. We still have, however, several problems to solve. The models discussed here are
in principle for static patients without movements i.e. intrafractional and/or interfractional movements.
The latter can probably be solved by daily imaging and adapting the today’s plan. This can be
accomplished by e.g. a library of plans or on-line re-optimisation and calculation. For intrafractional
movements, we have used PTV, ITV etc. and margin recipes to assure that the target get the correct
dose. This will in many cases lead to an over-irradiation of healthy tissues surrounding the tumour.
Techniques with synchronisation of the patient with the treatment delivery e.g. gating and tracking is
close to be clinical routine at many departments.
Another problem that has been noticed when introducing more accurate algorithms is a tendency
of optimizers to put in too much dose in the vicinity of the tumour to assure an adequate dose in the
whole tumour. This is especially noted when working with the PTV concept and tumours in low density
regions such as the lung. In these cases, one should probably use a combination of the type a and b
algorithms where optimisation is performed using type a and then followed by a recalculation with an
appropriate type b model. More physically one can describe this as an optimisation based on energy
fluence (pencil beam model) followed by dose calculation with full lateral electron modelling.

6. References
[1] Ahnesjö A and Aspradakis M A 1999 Phys. Med. Biol. 44 R99-155
[2] ICRU 1976 24 Bethesda, Maryland, USA
[3] Batho H F 1964 J. Canadian Ass. Radiologists 15 79-83
[4] Mackie T R et al 1985 Med. Phys. 12 188-96
[5] Lulu B A and Bjärngard B E 1982 Med. Phys. 9 372-7
[6] Lulu B A and Bjärngard B E 1982 Med. Phys. 9 907-9
[7] Sontag M R and Cunningham J R 1978 Radiology 129 787-94
[8] O'Connor J E 1957 Phys. Med. Biol. 1 352-69
[9] O'Connor J E 1984 Med. Phys. 11 678-80
[10] Wong J W and Henkelman R M 1983 Med. Phys. 10 199-208
[11] Mackie T R et al 1985 Med. Phys. 12 327-32
[12] Mohan R et al 1986 Med. Phys. 13 64-73

8
9th International Conference on 3D Radiation Dosimetry IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1234567890
847 (2017) 012037 doi:10.1088/1742-6596/847/1/012037

[13] Boyer A 1984 Med. Phys. 11 552-4

[14] Boyer A and Mok E 1985 Med. Phys. 12 169-77
[15] Fano U 1954 Radiat. Res. 1 237-40
[16] Ahnesjö A et al 1987 Acta. Oncol. 26 49-56
[17] Knöös T and Nilsson M 1987 Proc. 9th Int. Conf. on the Use of Comp. in Rad. Ther.
[18] Nilsson M and Knöös T 1992 Phys. Med. Biol. 37 69-83
[19] Ceberg C P et al 1996 Med. Phys. 23 505-11
[20] Storchi P and Woudstra E 1996 Phys. Med. Biol. 41 637-656
[21] Mohan R and Chui C S 1987 Med. Phys. 14 70-7
[22] Ahnesjö A et al 1992 Med. Phys. 19 263-73
[23] Knöös T et al 2006 Phys. Med. Biol. 51 5785-807
[24] Ulmer W et al 2005 Med. Biol. 50 1767-1790
[25] Van Esch A et al 2006 Med. Phys. 33 4130-48
[26] Tillikainen L et al 2008 Phys. Med. Biol. 53 3821-39
[27] Knöös T et al 1994 Phys. Med. Biol. 39 1609-28
[28] Knöös T et al 1995 Phys. Med. Biol. 40 1411-20
[29] Knöös T et al 1995 Radiother. Oncol. 37 74-80
[30] Weber L et al 1996 Med. Phys. 23 307-16
[31] Hurkmans C et al 1996 Phys. Med. Biol. 41 1277-90
[32] Hansson H et al 1999 Phys. Med. Biol.44 N201-7
[33] Ahnesjö A 1989 Med. Phys. 16 577-92
[34] Miften M et al 2000 Phys. Med. Biol. 45 817-33
[35] Nelson W R et al 1985 Stanford Linear Accelerator Report SLAC-265
[36] Fippel M 2004 Med. Phys. 31 1235-42
[37] Rogers D W 2006 Phys. Med. Biol. 51 R287-301
[38] Chetty I J et al 2007 Med. Phys. 34 4818-53
[39] Wareing T A et al 2001 Nucl. Sci. Eng. 138 256-68
[40] Gifford K A et al 2006 Phys. Med. Biol. 51 2253-65
[41] Vassiliev O N et al 2008 Int. J. Radiat. Oncol. 72 220-7
[42] Vassiliev O N et al 2010 Phys. Med. Biol. 55 581-98
[43] Fogliata A et al 2011 Radiat. Oncol. 6 82
[44] Fogliata A et al 2011 Med. Phys. 38 6228-37
[45] Hoffmann L et al 2012 Acta Oncologica 51 376-85
[46] Failla G A et al 2015 Varian white paper RAD 10156 10/10 350
[47] Ma C M and Li J 2011 Phys. Med. Biol. 56 3073-89
[48] Siebers J V et al 2000 Phys. Med. Biol. 45 983-95
[49] Mampuya W A et al 2016 J. Appl. Clin. Med. Phys. 8 6338
[50] Rana S et al 2014 Int. J. Cancer Ther. Oncol. 2 020421
[51] Mather M L et al 2003 Ultrasonics 41 551-9
[52] Baldock C et al 2010 Phys. Med. Biol. 55 R1-63
[53] Vial P et al 2006 Phys. Med. Biol. 51 5517-38
[54] Hurley C et al 2006 Nucl. Instrum. Meth. A 565 801-11
[55] Baldock C 2009 J. Phys.: Conf. Ser. 164 01200
[56] Hill B et al 2005 Med. Phys. 32 1589-97
[57] Hill R et al 2009 Med. Phys. 36 3971-81