ESTRO 36 Abstract Book

109 / 1096

S96 ESTRO 36
_______________________________________________________________________________________________
Fig. 1 demonstrated the capability of this QA system by
showing a source transit process and Fig. 2 indicated that
measured dwell time was affected by source separation.
Table 1(a) tabulated the measured dwell time for 3
different assigned dwell times with 5 mm separation
between source dwell positions. In all three scenarios, the
dwell time at starting position was close to the assigned
value. Dwell time at next dwell position experienced a
larger discrepancy up to 40% for 0.1 s dwell time. This
discrepancy in dwell time was due to the transit time for
which control computer could not fully account. Hence,
dwell time would be shorter than the assigned value
except at the starting position. Table 1(b) tabulated
measured dwell times at 3 different source separations
with 0.5 s assigned dwell time to assess the compensation
method stated. Discrepancy could be up to 0.33 s in 6 cm
separation. Transit time occupied a larger portion of the
dwell time for longer source separation.
Conclusion
Dwell time and transit time could be measured using the
fluorescent QA system with uncertainty down to 2 ms.
High temporal resolution in this system helped measure
the transit time accurately which could hardly be achieved
in commonly used QA systems. The effect of transit time
on actual source dwell time could be significant and was
not fully accounted for by treatment computer. Clinically
possible combinations, like 0.5 s dwell time and 5 mm
separation, could have a dosimetric error of 8%.
PV-0188 Improved class solutions for prostate
brachytherapy planning via evolutionary machine
learning
S.C. Maree1, P.A.N. Bosman2, Y. Niatsetski3, C. Koedood
er1, N. Van Wieringen 1, A. Bel1, B.R. Pieters1, T.
Alderliesten1
1 A cademic Medical Center, Radiation oncology,
Amsterdam, The Netherlands
2 Centrum Wiskunde & Informatica, Amsterdam, The
Netherlands
3Elekta, Veenendaal, The Netherlands
Purpose or Objective
In PDR and HDR prostate brachytherapy (BT), treatment
plans have to be created in a reasonably short time. In our
clinic, an initial p lan is automatically generated with an
optimization algorithm using a standard parameter set,
called a class solution (CS). Next, the plan is fine-tuned
manually using graphical optimization. The better the CS,
the less fine-tuning is required. We developed a method
to automatically find a CS such that the plans resulting
from the use of this CS match given reference plans as
good as possible, regardless of how these reference plans
were created.
Material and Methods
Twenty patients consecutively treated with PDR BT for
intermediate/high-risk prostate cancer were included.
Clinically acceptable reference plans were created in
Oncentra Brachy using manual graphical optimization
according to our clinical protocol.
To demonstrate our method, we learn CSs for Inverse
Planning Simulated Annealing (IPSA). Per organ, the IPSA
parameter set consists of an acceptable dose range and a
penalty value for violating this range. The ranges follow
from our clinical protocol, and the penalty values are
automatically learned for each patient individually (IPSA-
I) by minimizing the difference between the reference and
IPSA-generated plan using the evolutionary algorithm
known as AMaLGaM. Then, three CSs are compared:
- (CS-C) is the current clinical CS,
- (CS-M) results from a frequently used strategy for IPSA
by computing the mean of the IPSA-I parameters found
for the individual patients,
- (CS-S) is learned by using AMaLGaM again, but this time
aimed at minimizing the sum of plan differences for
multiple patients simultaneously.
Plan difference was measured by the root mean square of
the differences in selected DVH indices (Table). To
prevent overfitting, the data was randomly split into two
sets of 10 patients so that both CS-M and CS-S could be
learned twice: once on each half and validated on the
other half (2-fold cross validation).
Results
Our method is highly accurate when determining IPSA
parameters for individual patients (IPSA-I; dark purple
bars, Figure), with DVH indices of the reproduced plans
differing on average less than 2% of the reference plans
(Table). CS-S performs best for 13 of the patients, and has
the lowest average plan difference. CS-M has a larger plan
difference on average, but outperforms the current
clinical CS-C as well.
Conclusion
Our method for automatically determining class solutions
was found to be advantageous for our patient group,
outperforming the commonly used approach of taking the
mean of IPSA parameters. For individual patients, IPSA
parameters could automatically be found such that the
corresponding plans were very similar to the reference
plans. The performance gap between the latter and the
use of class solutions shows that there is still much room
for improvement by moving toward a patient-tailored
approach for automated BT planning. Our work achieves a
first step in that direction.