Professional Documents
Culture Documents
ORIGINALARBEIT
1. Introduction
$
Wissenschaftspreis der DGMP 2006 The advancement of dose delivery technology for in-
Tel.: +7071 2986055; Fax: +7071 294820. tensity modulated radiotherapy with photons or protons
E-mail: markus.alber@med.uni-tuebingen.de (IMRT, IMPT) has made numerical dose optimisation a
necessity. The price for the new freedom in dose shaping This paper is devoted to a specific class of L1 dose eva-
offered by these technologies has to be paid by a loss of luation models, their connection to biological dose-effect
transparency of the treatment planning process, the de- and normal tissue complication probability (NTCP) mo-
pendence on optimisation algorithms and the con- dels, and their efficacy both numerically and in controlling
sequential need to communicate the treatment intentions the dose distribution in normal tissues. These models are
to a computer. termed local dose-effect measures (LDEM). The field of
A very large variety of formulations of the radiotherapy L2 dose evaluation is of interest only to put the LDEM
optimisation problem with its inherent ambivalence be- dose evaluation in a general context. To start, a ‘‘standard
tween therapy success and harmful side-effects has been model’’ of biological dose optimisation is introduced with
conceived over the last 15 years (c.f. [1]). Despite this mul- the intent to provide a clean canvas for the following de-
titude, all these approaches serve the same purpose of ex- velopment. The LDEM class of models is a superset of the
pressing the clinical treatment intentions in a concise, nu- equivalent uniform dose formalism [8]. Its properties are
merically expedient, unambiguous and intuitive manner. discussed in the context of NT volume effects and associ-
Hence, it should not come as a surprise that some basic ated NTCP models. Each specific implementation of an
common traits can be identified. LDEM effects control over the dose distribution in a parti-
Optimisation equals evaluation and exploration. There cular manner, which is elucidated for two standard
are two distinct layers of evaluation of a therapeutic dose types of LDEMs in clinical applications. In an appendix
distribution. Radiotherapy commonly affects both target (Table 1), a table of typical model parameter values for a
volumes (TV) and normal tissues (NT) alike. The indivi- number of normal tissues is given.
dual effects in each NT should be mild or at least in a
justifiable relation to a gain in target effect. Once
quantified, effects may readily be prioritized in decision 2. Methods
making. However, for each NT, a value has to be as-
signed to a non-homogeneous dose distribution that 2.1. The standard model of biological optimisation and
reflects faithfully the intuitive, implicit rules of clinical ex- local dose-effect measures
perience. Henceforth, dose evaluation on a per-organ ba-
sis with a focus on biological effectiveness shall be termed For dose optimisation, high- and low-dose volumes in the
level one (L1), while the search for the right balance be- same normal tissue or target volume need to be balanced
tween a set of TV and NT effects shall be termed level two (L1), and the total effects of various NT and TV need to
(L2) evaluation. be weighed against each other (L2). In other words, there
Despite the high degree of non-explicit clinical know- have to be mathematical functions that assess the meaning
ledge and the difficulty to extract solid data from clinical of dose in a subvolume relative to the total volume of one
studies, L1 dose evaluation is best left to the computer organ, and other functions that poise the total effect of the
and taken out of the hands of the clinician. The dose dis- dose distribution in one organ against the others, inclu-
tribution in a NT volume is a massive amount of ding the targets.
data, even if compacted into a dose volume histogram It has been suggested to express the total effect in a
(DVH). Further, it is usually impossible to alter one TV(NT), denoted as F (G), by a sum over all volume ele-
aspect of a dose distribution without affecting it as a ments belonging to the volume in question [9,10]
whole, including all other NT and TV. From this inabi- 1X N
lity to control a dose distribution in all its facets, the need F¼ f ðDi Þ, ð1Þ
N i¼1
for dose-effect models arises. These models serve both as
the knobs and levers to steer the dose distribution into a 1X N
fast dose evaluation- and optimisation algorithms and Once a solution is found for a particular set of Lagrange
makes the effect density practically independent of the multipliers, these may be altered incrementally to navigate
voxel size of the chosen patient model. The limitations on the Pareto frontier to other, more preferable solutions
with respect to long-range, non-local biological effects are with a different balance of NT and TV effects.
discussed in the final section of this paper. Naturally, there is a tight antagonistic coupling be-
Having obtained the total effects of n target volumes tween NT effects and TV effects. If no NT effect was do-
and m normal tissues, a particular dose distribution can be se limiting to the targets, the solution to the optimisation
quantified by the vector problem would be trivial. Hence, it is also possible to ex-
press the best TV effects as a function of the NT effects.
ðF 1 ; . . . F n ; G 1 ; . . . ; Gm Þ 2 ½0; 1½nþm . (3) This results in a different parametrization of the Pareto
frontier in terms of the NT effects. In practice, one could
Clearly, it is not possible to minimize all of these effects choose li ¼ 1; i ¼ 1; . . . ; n, assuming that all target vol-
simultaneously or else the problem would have the trivial ume effects are normalized identically, and then minimize
solution ðF 1 ; . . . F n ; G1 ; . . . ; G m Þ ¼ 0. Obviously, some the sum of all target effects such that Gi pC i ; i ¼ 1 . . . m.
combinations of effects cannot be obtained from physical The target effects can be controlled by altering the con-
dose distributions. This reflects the practical experience straint limits C i . This constraint optimisation approach
that improving the dose distribution in one volume usu- requires a more sophisticated optimisation algorithm [4],
ally makes it worse in others. but offers support for the navigation on the Pareto fron-
The physics of dose deposition introduce a boundary in tier [7]. A further exploration of this subject of L2 evalua-
the solution space that separates feasible from unphysical tion is beyond the scope of this paper, yet the relationship
dose distributions. The optimum solution for a given pa- between apparently incongruous approaches to L2 op-
tient, expressed in its particular combination of achievable timisation may lend some generality to the following de-
F and G values, has to be found on this boundary. Since velopment, which occurs in the framework of Eq. (4).
greater values of F or G mean worse dose distribution
properties, any dose distribution whose total effect values
lie above the boundary are not optimum, while physical 2.2. The volume effect and equivalent uniform dose
dose distributions cannot lie below.
In order to arrive at a point on this boundary manifold, Although the form of Eqs. (1), (2) can be motivated by a
the standard approach is to identify the effect functions F, mean-field approximation of non-local biological effects,
G as cost functions of an optimisation problem, and com- the true benefit of this formalism lies in the transparent
bine them to a Lagrange function L by virtue of and effective way by which these effect functions offer
control over the shape of the dose-volume histogram of a
X
n nX
þm
L¼ li F i þ li G in , (4) NT volume. Starting from the notion that the local effect
i¼1 i¼nþ1 densities g reflect the damage wrought in a volume ele-
P ment of some normal tissue by the applied dose, it is clear
with Lagrange multipliers li 2 ½0; 1, nþm i¼1 li ¼ 1. By con- that g is a continuous, monotonically increasing function
vention, the optimum dose is obtained by minimizing L. of dose. It is also obvious that g is bounded from above as
By varying the values of ðli Þ1::nþm , any possible combina- the maximum damage in a tissue is limited by its total ob-
tion of total effects can be obtained. Under very general literation. So in general, g is of a sigmoidal shape. How-
assumptions and for most practical purposes, there exists ever, this upper bound to the NT effect may not become
a one-to-one relationship between the n þ m 1 dimen- apparent in clinical dose regions.
sional hyperprism of possible ðli Þ1::nþm vectors and the The salient feature of g is the representation of the typi-
entire solution space of the problem. Romeijn [12] has cal volume effect of a normal tissue dose response, i.e. its
shown that virtually all so far suggested approaches to L2 deviation from a step function (no volume effect). The
evaluation can be reduced to the form of Eq. (4). All spe- volume effect details how much tissue volume has to be
cializations result in a mapping of the multiplier space spared in order to gain some added tolerance in a high-
onto itself, which may be numerically expedient, but does dose subvolume. With a small volume effect, the return on
not alter the solution space or the optimum solution. volume sparing is low. The ideal DVH in this case is a
Recently, the terminology of Pareto optimality at- steep drop at the tolerance dose. This kind of behaviour is
tracted attention in this context [5,6]. The aforementioned usually preferred for organs showing a serial, chain-like
boundary is identical to the Pareto frontier P and can be complication mechanism: just as the chain breaks when
obtained in this framework as a manifold in the n þ m di- one link gives way, the complication occurs if a small
mensional solution space parametrized by ðli Þ1::nþm subvolume of the organ is destroyed. In contrast, with a
large volume effect small volumes may receive very high
PðF ; GÞ ¼ ðF 1 ; . . . F n ; G 1 ; . . . ; G m Þðl1 ; . . . ; lnþm Þ. (5) doses if at the same time large volumes are effectively
ARTICLE IN PRESS
spared. The ideal DVH in this case shows a steep drop at the tissue is acceptable. Hence, the clinical endpoint is not
low to intermediate doses, and levels off to a shallow tail tied to the integrity of the entire organ as in serial
towards high doses. This is often associated with a paral- complications, but rather to the integrity of a critical
lel, rope-like complication mechanism: a fraction of the subvolume of the organ. Requiring the entire sigmoidal
strands may tear, yet the rope may still hold. shape of the local effect density is the hallmark of parallel
In mathematical terms, the volume effect answers the complication mechanisms. Thus, the full sigmoidal dose
question: if the homogeneous irradiation of a fractional effect needs to be described. For the gEUD model, this
volume v with a dose D results in a certain level of toxicity, extension to the full range results in g ¼ 1=ð1 þ ðD0 =DÞk Þ
how does a change of fractional volume dv alter the dose (pEUD), the reconstruction unit model yields g ¼ 1=ð1 þ
dD at the same toxicity? This is captured in the function expðsðD D0 ÞÞÞ (pRU) [15]. The reference dose para-
Qðv; DÞ40 in the following differential equation meter D0 marks the point of inflexion.
dv ¼ Qðv; DÞdD. (6)
It is easy to see, for G ¼ const., that 2.3. Serial complication mechanisms
vg0
Q¼ (7)
g For typical parameters, e.g. k ¼ 12; s ¼ 0:15, the gEUD
in the local effect formalism. By virtue of this relationship, and sRU models show a steeply increasing, strongly cur-
it is straightforward to construct an efficient cost function ved local effect density, and an associated small volume
for optimisation from NTCP models. At the core of any effect. For example, a dose increase by 10 per cent ne-
relevant NTCP model lies a representation of the volume cessitates a reduction of irradiated volume by a factor 0:3.
effect, which can be laid bare by Eq. (6). One NTCP mo- This strongly non-linear behaviour has two consequences.
del has recently attracted interest in the context of dose Firstly, the total effect will always be dominated by the
optimisation. The particular choice of gðDÞ ¼ Dk ; kX1 volume elements receiving the highest doses. This effect is
leads to the generalized equivalent dose model (gEUD) so strong that volume elements with less than 80 per cent
[8] which is the core of the well established Lyman-Kutc- of the maximum dose in this organ contribute virtually
her-Burman model [13,14]. nothing to the total effect. Secondly, the cost of increasing
Under the above conditions on g, it is always possible to the dose to an already hot subvolume rises tremen-
form duously. Since by virtue of Eq. (4) all effects are con-
nected, this rise has to be compensated by a drop in
EUD ¼ g1 ðGÞ. (8) one of the effect functions of other NT, or more likely, a
The EUD represents the dose that causes the same effect if target volume. This ensures that both the risk is spread
applied homogeneously to the entire organ volume. The equally among NT, and that cold spots in a TV are tightly
advantage of expressing the effect in terms of a dose is that coupled to hot spots in a NT. In consequence, these low-
the numerical value becomes more meaningful, and can be volume-effect cost functions act strictly dose limiting by
related to clinical experience. themselves, which makes the adjustment of the associated
It has been shown (c.f [15,16]) that the gEUD - power Lagrange multiplier less critical. As a corollary, the
law model provides a good approximation to the ob- power-law dependence of the total effect on basically the
servable volume effect of serial complications, where the maximum dose may even serve as a guide to adjust the
volume effect is small. For optimisation algorithm pur- magnitude of Lagrange multipliers should the total effect
poses, it provides the additional benefit that it is dose- of this NT be too high [11].
scale invariant [17]. However, it appears as if the power This discussion already emphasizes the aspect of con-
law arises only as an approximation to a more fundamen- trolling the shape of the DVH by the choice of local effect
tal dose-response in the limit of low doses/low complica- density. The dose distribution of each NT and TV is partly
tion rates [15]. The serial reconstruction unit (sRU) model a consequence of the chosen local effect density and the
suggests an exponential local effect density g ¼ expðsDÞ, interplay of different NT and TV, mediated by patient
with s being typically in the range of 0:15 Gy1 . geometry and the physics of dose deposition. Total con-
It may appear as a contradiction to the previous state- trol of the dose distribution in one volume usually means
ment about the boundedness of g that both presented loss of control in all other volumes. Hence, it is important
effect densities tend to 1 with dose. This is a consequence to realize which aspects of a dose distribution are control-
of the fact that for most serial complications, the point of led by a particular effect density function, and to which
inflexion of the sigmoid is well above the clinical dose aspects this function is blind.
range so that these functions still provide a faithful ap- During dose optimisation, each cost function will
proximation for the convex branch of the sigmoidal dose reward a redistribution of dose such that the total effect
response [15]. If clinical doses reach the saturation level of is diminished. Hence, it is illustrative to look at the drop of
the effect density, this means that a partial obliteration of local effect density per dose level, dg=dD. In Figure 1 the
ARTICLE IN PRESS
3. Results
2.4. Parallel complication mechanisms
These LDEM models may also be defined for targets, e.g.
Applying the same reasoning to the full sigmoidal local by means of f ¼ expðaDÞ and can be combined with
effect density for a parallel complication mechanism, it physical cost functions such as quadratic penalties, which
obtains that the derivative dg=dD is peaked around the fit into the same formalism. The routine IMRT treatments
point of inflexion D0 . Figure 2 shows the corresponding at University of Tübingen have made use of this for-
weight control arrows in a typical DVH. It becomes appa- malism since September 2001, and to date in excess of
rent that this effect density type acts like a watershed. As 800 patients have been treated with dose distributions op-
the gain of reducing the dose in volume elements close to timized on these principles. In combination with the use of
D0 is greatest, the dose is redistributed towards the extre- hard constraints for normal tissue objectives, this has
mes, where dose does not mean much harm yet or cannot resulted in very homogenous dose distributions in terms
cause any more damage than already wrought. In con- of EUD across the patient populations of various case
ARTICLE IN PRESS
The proposed LDEM is primarily geared towards this can be assessed by computing the EUD for a set of already
transparency and numerical expediency. At the same time, applied dose distributions. In the end, nothing can replace
it is possible to express a wide range of normal tissue vol- good judgement, for which it is helpful to know which
ume effects with these models. For dose optimisation, the parts of the dose distribution are controlled by the
normal tissue volume effect is arguably the most impor- LDEM.
tant property of any organ at risk, and is of paramount The cost for the detailed consideration of the volume
importance if dose escalation is to be performed safely. effect is that LDEM are non-linear functions which re-
The LDEM compute a single number that represents the quire more sophisticated optimisation algorithms than
total effect of a dose distribution in the volume to which e.g. linear cost functions. Of particular importance is the
the model is applied. This number is a ‘‘biological’’ quan- existence of locally optimum solutions. In general, it can
tity in that it considers the volume effect. It is possible to be shown that if the local effect density is convex (curva-
map the total effect onto a normal tissue complication ture like an upturned parabola), local minima cannot
probability with a one-to-one relation, so that any LDEM occur. If the local effect density has a sigmoidal shape,
can be incorporated in a NTCP model. Vice versa, almost local minima cannot be excluded, but this is no guarantee
all NTCP models can be expressed in terms of a LDEM. It that they do occur. Their occurrence is frequently preven-
follows that even approaches like the optimisation of the ted by the physics of dose deposition which does not per-
probability of complication-free tumour control Pþ are mit to change the dose distribution in isolated subvolumes
only a special case of Eq. (4). only. Further, the smaller the curvature of the local effect
The fundamental approximation and limitation of the density, the less likely is the existence of local minima.
LDEM is the reliance on local effect densities, which can In consequence, the pEUD/pRU models behave better
constitute a problem if non-local biological effects played numerically than exact dose-volume constraints, and
a role. Evidence of such effects can be found in animal allow some more flexibility in finding the optimum dose.
experiments with ultra-small fields and very high doses It is safe to assume that for all practical purposes, the
[18]. The reconstruction unit models introduced above as Lagrange function is convex so that the statements made
LDEMs have been extended to take into account non-lo- about the existence of a manifold of optimum solutions
cal tissue repair effects [15]. In the context of radiotherapy as the boundary between physical and unphysical dose
with photons, electrons and protons, the penumbra of the distributions hold true.
dose distributions is usually so shallow that non-local re- The LDEM formalism encorporates biological model
pair effects do not play a role and do not constitute a li- assumptions, which will eventually be verified by clinical
mitation to the LDEM approach. Heavy ions may indeed studies. Its primary purpose is to provide a toolbox for
be an exception. In these cases, the LDEM would ove- steering the dose optimisation. As with all tools, their
restimate the damage to normal tissue, and consequenti- sheer existence does not prove their utility. However, their
ally err to the safe side. design was guided by the intent to put dose optimisation
A practical limitation of LDEM is the undeniable spar- on a formal basis and to make control over the optimisa-
city of clinical data to pinpoint the parameters required by tion process more transparent, which are arguably worth-
the model. For some organs, like rectum and lung, good while goals.
parameters sets exists. For others, where clinical compli-
cations are sparse, parameters have to be tried and tested
in planning studies and clinical routine. The parameters in
the appendix have been distilled from 6 years of clinical Acknowledgements
application of the LDEM in Tübingen, but are by no
means final (Table 1). When in doubt, it is always prefer- The author was awarded the Research Prize of the Ger-
able to err to the safe side, which in the case of serial com- man Association of Medical Physicists (DGMP) in 2006
plications means to assume a smaller volume effect, i.e. for his research in the field of radiotherapy dose opti-
greater k or s. The range of clinically applied total effects misation.
ARTICLE IN PRESS
Table 1
Effect density types and parameters for selected organs/complications. Parameters are partially obtained from clinical data, partially
from planning studies. EUD and critical volume constraints should be lower than given values if possible.
of prostate cancer using EUD- and dose-volume based NTCP mo- [18] Bijl HP, van Luijk P, Coppes RP, Schippers JM, Konings AWT,
dels. Int J Radiat Oncol Biol Phys 2007;67(4):1066–73. van der Kogel AJ. Unexpected changes of rat cervical spinal cord
[17] Alber M, Nüsslin F. A representation of a NTCP function for local tolerance caused by inhomogeneous dose distributions. Int J Radiat
complication mechanisms. Phys Med Biol 2001;46:439–47. Oncol Biol Phys 2003;57(1):274–81.