Advanced

Automation of robust Pareto front-based radiotherapy treatment planning for prostate cancer patients

Eliasson, Niklas (2019) MSFT01 20181
Medical Physics Programme
Abstract
Purpose/objective
The main objective was to create a software that automates the process of creating treatment
plans used for Pareto front-based dose planning for prostate cancer patients. A second
objective/purpose was to add a robustness test to this program to evaluate the effect of prostate
movements on the treatment plans.

Material/method
An IronPython program was designed to control and collect information from the treatment
planning system (TPS) RayStation 5 and by using built-in libraries made by the creators of the
Raystation 5.
The patients selected were prostate cancer patients with treatment of only the prostate, not
including nearby lymph nodes or seminal vesicles. Hypo-fractionation treatment plans were
created... (More)
Purpose/objective
The main objective was to create a software that automates the process of creating treatment
plans used for Pareto front-based dose planning for prostate cancer patients. A second
objective/purpose was to add a robustness test to this program to evaluate the effect of prostate
movements on the treatment plans.

Material/method
An IronPython program was designed to control and collect information from the treatment
planning system (TPS) RayStation 5 and by using built-in libraries made by the creators of the
Raystation 5.
The patients selected were prostate cancer patients with treatment of only the prostate, not
including nearby lymph nodes or seminal vesicles. Hypo-fractionation treatment plans were
created with a prescription of seven fractions of 6.1Gy with a total dose of 42.7Gy, one fraction
every other day.
A comparison was made between a treatment plan created by a dose planner and Pareto fronts
extracted from 1440 treatment plans automatically generated with the plan generation program.
The robustness test was evaluated on one patient by using an isocenter shift of (x, y, z) =
(0.45cm, 0.02cm, -0.58cm).

Result
The software consisted of two programs. The first program used the optimizer and dose
calculator in RayStation 5 to create deliverable treatment plans. It inputted different objective
functions and/or weights to the optimizer. For each change, the optimizer would find a new
optimal treatment plan. It saved DVH-data for evaluation of the plans. A built-in robustness
test was added to the program to test the effect of prostate movements on the treatment plans it
constructed. It moved the finished treatment plan’s isocenter and the dose difference was
calculated.
The second program, created for evaluation, loaded the saved data from the program generating
the treatment plans. In the program, multiple data sets was loaded and compared. It visualized
both the Pareto fronts based on the collected plans and all the dose volume histograms (DVHs)
for one plan at a time (for a selected plan from the first graph). The program could determine if
treatment plans were Pareto optimal and/or clinically acceptable. It was also used to visualize
the robustness test, where the static treatment plans and treatment plans with a moved isocenter
were plotted in the same graph.
The programs automated the process and reduced the work needed to only some preparation
before starting the program. The time to create treatment plans for a Pareto front was greatly
reduced, as the program could save one plan every 2-3 minutes.
When comparing a Pareto front consisting of automatically generated treatment plans and a
plan created by a planner, the dose planner’s treatment plan ended up near the Pareto front in
all cases.

Discussion
Prostate cancer was selected for this study due to the fact that it is a comparably simple case
involving only a few OARs. The only trade-off that needed to be visualized is the one between
the target and the rectum. Thus, only the rectum goal needs to be changed to be able to show
the trade-off. If more OARs would be added, it would have taken longer time to generate
enough treatment plans to represent the trade-offs.
There are, however, improvements that could be made. For further automation, and to decrease
the generation time, machine learning could be used.
The robustness test was able to show how the dose distribution would be affected by an
isocenter movement. Several improvements could be made. For instance, there are several
fractions in a treatment course, and the same movement does not occur each time. There might
be a continuous movement during the treatment delivery, not only one big movement.
Furthermore, the data used was for a 30-minute interval, while in reality, a patient has come
and gone in less than half that time.

Conclusion
A program has been developed and implemented that can be used for automation of creation of
treatment plans used for Pareto front-based dose planning. A robustness test was built-in to
allow for comparison between the created plans with respect to how much prostate motion
would affect them. Features such as machine learning would be a good tool to further automate
the process and to reduce the generation time. (Less)
Popular Abstract (Swedish)
Antalet rapporterade cancerfall i Sverige har mer än dubblerats sedan 1970-talet. Cancerfonden rapporterade att 61100 personer blev diagnostiserade med cancer under 2015 och att prostatacancer är den mest förekommande typen av cancer. Ett vanligt sätt att behandla prostatacancer är extern strålbehandling, där en strålkälla (vanligtvis en linjäraccelerator) som befinner sig utanför kroppen används för att bestråla tumören med joniserande strålning.

För varje patient skapas en individuell behandlingsplan, vilken innehåller instruktioner för hur behandlingsmaskinen skall leverera behandlingen. Då planen skapas definieras mål för hur mycket olika områden skall eller får bestrålas. Syftet för målen är att säkerställa en bra behandling... (More)
Antalet rapporterade cancerfall i Sverige har mer än dubblerats sedan 1970-talet. Cancerfonden rapporterade att 61100 personer blev diagnostiserade med cancer under 2015 och att prostatacancer är den mest förekommande typen av cancer. Ett vanligt sätt att behandla prostatacancer är extern strålbehandling, där en strålkälla (vanligtvis en linjäraccelerator) som befinner sig utanför kroppen används för att bestråla tumören med joniserande strålning.

För varje patient skapas en individuell behandlingsplan, vilken innehåller instruktioner för hur behandlingsmaskinen skall leverera behandlingen. Då planen skapas definieras mål för hur mycket olika områden skall eller får bestrålas. Syftet för målen är att säkerställa en bra behandling samtidigt som närliggande organ skyddas. Till exempel bör tumören bestrålas så mycket som läkarna har ordinerat samtidigt som bestrålningen av känsligare organ som till exempel ändtarmen hålls låg. Genom att bestråla känsliga områden så lite som möjligt och stäva efter att minst uppnå de kliniska målen så minskas risken för de negativa konsekvenser som är kopplade till att bestråla området. De satta målen konkurrerar ofta med varandra och planerarna får finna en balans mellan rätt bestrålning av tumören och så lite bestrålning som möjligt till de känsligare områdena i närheten av tumören.

För de mer avancerade behandlingsmetoderna används inversdosplanering för att skapa en behandlingsplan. Då ger dosplaneraren sina mål för fallet till ett datorprogram som används för att skapa behandlingsplanen. Programmet använder en optimeringsalgoritm för att finna en optimal behandlingsplan baserat på de mål som satts och viktningen mellan målen. För att påverka vilken optimal behandlingsplan som beräknas fram av programmet så modifierar dosplaneraren de givna målen och hur viktiga de ska vara.

Det är vanligt att målen konkurrerar med varandra vilket kräver att en balans mellan målen hittas. För att finna de bästa möjliga behandlingsplanerna kan man söka efter Pareto-optimala planer. En behandlingsplan är Pareto-optimal om det inte går att förbättra ett av målen utan att försämra ett av de andra målen. Tillsammans kallas alla Pareto-optimala planerna en Pareto- font. Planerna i en Pareto-fronten är inte optimal för alla mål utan de planer som är bättre för ett mål blir sämre för ett annat mål.

Dessvärre är det tidskrävande att skapa och samla in information om många behandlingsplaner, men hela processen går att automatisera med hjälp av programmering. Ett program som efterliknar mänsklig planerare och följer en inversdosplaneringprocess kan skapas. Programmet använder därefter information från de skapade behandlingsplanerna för att bestämma vilka av dem som är Pareto-optimala i förhållande till de satta målen. Därefter kan en front av behandlingsplaner visas upp för att få en bättre överblick över konkurrensen mellan målen.

Skillnader mellan behandlingstillfället och tillfället för skapandet av behandlingsplanen kan uppstå på grund av prostatans rörelse i kroppen. Detta påverkar behandlingen, som då inte utförs som planerat. Genom att undvika behandlingsplaner som påverkas mycket av rörelsen så kan rörelsens påverkan minimeras. Hur mycket olika behandlingsplaner påverkas kan testas genom att efterlikna rörelsen i datorn då behandlingsplanen skapas och redan då utesluta planer som påverkas för mycket av prostatans rörelse.

Automatisering av dosplaneringsprocessen ger många fördelar. Arbetet som krävs för att utvärdera behandlingsplaner med en front av Pareto-optimala planer blir mindre. De skapade behandlingsplanerna kommer alltid få samma kvalité, vilket inte påverkas av vilken dosplanerare som skapar behandlingsplanen eller hur lång tid dosplaneraren lägger på att skapa behandlingsplanen. Dessutom kan tillägg skapas till programmet vilket ger möjligheten att testa behandlingsplanerna för påverkan av prostatans rörelse, eller lägga till andra tillägg som maskininlärning vilket hjälper programmet finna de bättre lösningarna snabbare. (Less)
Please use this url to cite or link to this publication:
author
Eliasson, Niklas
supervisor
organization
course
MSFT01 20181
year
type
H2 - Master's Degree (Two Years)
subject
language
English
id
8999113
date added to LUP
2020-01-07 09:32:50
date last changed
2020-01-07 09:32:50
@misc{8999113,
  abstract     = {Purpose/objective
The main objective was to create a software that automates the process of creating treatment
plans used for Pareto front-based dose planning for prostate cancer patients. A second
objective/purpose was to add a robustness test to this program to evaluate the effect of prostate
movements on the treatment plans.

Material/method
An IronPython program was designed to control and collect information from the treatment
planning system (TPS) RayStation 5 and by using built-in libraries made by the creators of the
Raystation 5.
The patients selected were prostate cancer patients with treatment of only the prostate, not
including nearby lymph nodes or seminal vesicles. Hypo-fractionation treatment plans were
created with a prescription of seven fractions of 6.1Gy with a total dose of 42.7Gy, one fraction
every other day.
A comparison was made between a treatment plan created by a dose planner and Pareto fronts
extracted from 1440 treatment plans automatically generated with the plan generation program.
The robustness test was evaluated on one patient by using an isocenter shift of (x, y, z) =
(0.45cm, 0.02cm, -0.58cm).

Result
The software consisted of two programs. The first program used the optimizer and dose
calculator in RayStation 5 to create deliverable treatment plans. It inputted different objective
functions and/or weights to the optimizer. For each change, the optimizer would find a new
optimal treatment plan. It saved DVH-data for evaluation of the plans. A built-in robustness
test was added to the program to test the effect of prostate movements on the treatment plans it
constructed. It moved the finished treatment plan’s isocenter and the dose difference was
calculated.
The second program, created for evaluation, loaded the saved data from the program generating
the treatment plans. In the program, multiple data sets was loaded and compared. It visualized
both the Pareto fronts based on the collected plans and all the dose volume histograms (DVHs)
for one plan at a time (for a selected plan from the first graph). The program could determine if
treatment plans were Pareto optimal and/or clinically acceptable. It was also used to visualize
the robustness test, where the static treatment plans and treatment plans with a moved isocenter
were plotted in the same graph.
The programs automated the process and reduced the work needed to only some preparation
before starting the program. The time to create treatment plans for a Pareto front was greatly
reduced, as the program could save one plan every 2-3 minutes.
When comparing a Pareto front consisting of automatically generated treatment plans and a
plan created by a planner, the dose planner’s treatment plan ended up near the Pareto front in
all cases.

Discussion
Prostate cancer was selected for this study due to the fact that it is a comparably simple case
involving only a few OARs. The only trade-off that needed to be visualized is the one between
the target and the rectum. Thus, only the rectum goal needs to be changed to be able to show
the trade-off. If more OARs would be added, it would have taken longer time to generate
enough treatment plans to represent the trade-offs.
There are, however, improvements that could be made. For further automation, and to decrease
the generation time, machine learning could be used.
The robustness test was able to show how the dose distribution would be affected by an
isocenter movement. Several improvements could be made. For instance, there are several
fractions in a treatment course, and the same movement does not occur each time. There might
be a continuous movement during the treatment delivery, not only one big movement.
Furthermore, the data used was for a 30-minute interval, while in reality, a patient has come
and gone in less than half that time.

Conclusion
A program has been developed and implemented that can be used for automation of creation of
treatment plans used for Pareto front-based dose planning. A robustness test was built-in to
allow for comparison between the created plans with respect to how much prostate motion
would affect them. Features such as machine learning would be a good tool to further automate
the process and to reduce the generation time.},
  author       = {Eliasson, Niklas},
  language     = {eng},
  note         = {Student Paper},
  title        = {Automation of robust Pareto front-based radiotherapy treatment planning for prostate cancer patients},
  year         = {2019},
}