AAPM2015 - Statistical Learning to Predict MLC Errors

Post on 22-Jan-2018

773 views 2 download

Transcript of AAPM2015 - Statistical Learning to Predict MLC Errors

A Statistical Learning Approach to the Accurate Predictionof MLC Errors During VMAT DeliveryJoel Carlson

Jong Min Park

So-Yeon Park

Jong In Park

Yunseok Choi

Sung-Joon Ye

MLCs move in complex ways

Prostate Plan: Low complexity H&N plan: High complexity

** ~50x speed **

Complex movements lead to errors in leaf positions

How can we quantify these errors? Planned Delivered

Goal:Create a more realistic representation in the TPS of

where the MLC leaves will be upon delivery

How can we predict these errors?

How do these errors impact dose delivery accuracy?

How can we quantify these errors?

How can we predict these errors?

How do these errors impact dose delivery accuracy?

Quantifying the difference between planning and delivery74 H&N or Prostate VMAT plans from 3 institutions

Dicom RT

Planned Positions

DYNALOG

Delivered Positions

Errors(Prediction Target)

Difference

How can we quantify these errors?

How can we predict these errors?

How do these errors impact dose delivery accuracy?

Goal:Create a more realistic representation in the TPS of

where the MLC leaves will be upon delivery

We first extracted a rich feature set from the DICOM-RT plan files

Using only information available before plan delivery

~150 features* quantifying the MLC leaf motion

*list available, just ask!

Using a validation set we chose the best features

All Features

Build Model

Vary Features

Error* Minimized?

Predictions

NoFinal Model

Yes

Predictions

Report Statistics

Training Plans

(N = 3)Validation Plans

(N = 6)

Testing Plans

(N = 65)

*Root Mean Squared Error

Results:

Cubist (Decision Tree)

Best performing algorithm

Best performing feature set:

Velocity, Position, Direction, Movement Category, Bank

Results: The errors are well predicted by the machine learning algorithms

Results: Visualizing the movement of a single MLC leaf

~3.5mm

Results: Visualizing the movement of a single MLC leaf

How can we quantify these errors?

How can we predict these errors?

How do these errors impact dose delivery accuracy?

Goal:Create a more realistic representation in the TPS of

where the MLC leaves will be upon delivery

Calculate the gamma pass rates

DeliveredPlanned

Predicted Delivered

Eclipse Trilogy + MapCheck2

?

?

Passing rates are improved by using predicted positions

There exist errors between planned and delivered MLC positions

These errors are predictable at the planning stage

Utilizing predicted positions:• Increases gamma passing rates

• Leads to a more realistic representation of where the leaves will be upon delivery

In conclusion

Explore differences in patient DVHs• In progress

Integrate predictions into TPS• Will give planners a better view of what will be

delivered

Publish fully reproducible code and data

Future work

A Statistical Learning Approach to the Accurate Predictionof MLC Errors During VMAT DeliveryJoel Carlson

Jong Min Park

So-Yeon Park

Jong In Park

Yunseok Choi

Sung-Joon Ye

Thank you for listening!

Slides Answering Potential Questions

The following slides serve as supplemental material for answering audience questions

Planned

Predicted

SMG (R)

Parotid (L) Parotid (R)

PTV_67.5

PTV_54

SMG (L)

PTV_48

• Numerical Values:• Error Magnitude

• MLC Index

• Width and Mass of leaf

• Positions• ±5 CPs

• ±5 CPs of both adjacent MLCs

• Velocities• ±5 CPs

• ±5 CPs of both adjacent MLCs

• Accelerations• ±5 CPs

• ±5 CPs of both adjacent MLCs

• Momentum

All Features• Categorical

• Whether the MLC was previously at rest, coming to a stop, moving before and after, single CP movement

• Whether adjacent MLCs were both moving in the same direction, both opposite, same/opposite, or at rest

• Moving towards (push) or away (pull) from the isocenter

• The CP at which the error occurred

Cubist• “…is a rule-based

model where a tree is grown, and each of the terminal leaves contain regression models. These models are based on the predictors in previous splits.”