TRiP98 algorithms
Algorithms
Follows a short description of some algorithms used in
TRiP98
.
Its purpose is
to give a basic understanding, for a more detailed
description please look up the appropriate publications.
-
cl
, short for classic
This is the default algorithm.
It is based on the coordinate grid of the raster scanner
positions plus a water equivalent z-coordinate to account for
the depth. The width of each beam spot is accounted for
by distributing its intensity across its neighbours,
thereby establishing modified intensities in each raster position.
Dose calculation on the CT grid is performed by locating for each voxel its
4 neighbouring raster points and interpolating the modified intensities.
This method is fast and reasonably accurate.
It should, however, not be used for multifield optimization,
and does not give good results when CT and raster grid have
very different spacings.
-
ap
, short for allpoints
This method takes into account explicitly all neighbouring
raster beam spots which may contribute to a given voxel.
The Gaussian shape of each beam spot is explicitly calculated
(pencil beam).
This method is more accurate than the classic
algorithm, but takes considerably longer time.
It should be used e.g. for multifield optimization or
irregular raster patterns.
-
ms
, short for multiplescatter
This method works similar as allpoints
,
but takes into account the broadening of each beam spot
as a function of depth. It is more accurate than allpoints
,
but takes more time. At least for carbon it should be used only
for calculations at large depths.
multiplescatter
requires
DDDs
which contain extra columns describing the beam broadening,
otherwise the results will be as for allpoints
.
For ms
and ap
,
a divergence of the beam is considered if nonzero
scannerxdistance(),scannerydistance()
values are specified.
No divergence is considered for cl
and,
for the time being, for
Beam's Eye-View
calculations.
-
cl
, short for classic
This is the default, based on the original implementation
of the prescriptions of the LEM model.
Because internally random sampling of particle spectra
is used, the results show statistical fluctuations.
-
ld
, short for lowdose
This is still according to the LEM model, but uses
approximations for therapeutic dose levels,
allowing to speedup spectral interpolations by about an
order of magnitude or more.
In addition, results are no longer subject to statistical
fluctuations.
lowdose
may be used for time consuming calculations,
e.g. multifield optimization.
-
H2Obased
This is the classical default used for single field optimization.
It uses an optimization grid based on a water-equivalent
grid in beams-eye-view, with simplifications allowing
fast optimization. It does not, however, include lateral extension of
the beam spots, i.e. it always uses the classic
dose algorithm
and cannot perform simultaneous multifield optimization.
In the first step of this optimization a simple algorithm
is used to match the depth dose profiles ("Bragg peaks") with the
desired dose distribution. The second steps invokes gradient based
search algorithms, which from version 1001d onwards are the same as for
CTbased
.
H2Obased
by default is used to provide start values for
the CTbased
optimization.
-
CTbased
This must be used for simultaneous multiple field optimization,
but may also be selected for a single field.
It is based on the grid of the given CT, and considers the given
target VOI plus OARs. It allows considerably better optimization,
but takes a lot more iterations (several hundred) than
H2Obased
and eats a lot more memory
(several GBytes).
At present four algorithms can be used for optimization:
-
bf
, short for bortfeld
-
gr
, short for gradient
-
cg
, short for cjggrad
(default, for historical reasons)
-
fr
, short for fletcher-reeves
The algorithms give similar results,
but cg
usually converges faster (by about a factor 2 to 3)
than bf
whereas bf
converges "smoother".
gr
was introduced later,
it is the faster than cg
and also
more stable, so it should be preferred over this one.
As was found out recently, fr
converges yet faster than gr
by some factor four (for the same result), so if you're in a hurry,
take this one.
Last updated:
$Id: trip98alg.html,v 1.7 2012/08/16 17:13:55 kraemer Exp $
M.Kraemer@gsi.de