26/07/2018

1

Practical Examples of AutoPlanningVia Advanced Scripting

Ben Archibald-Heeren30/07/2018

Conflicts of Interest

No relevant financial relationships to disclose

ICON group has a development working relationship with Varian Medical and Breathwelldevices

Personal non-financial affiliation with the University of Wollongong

The above working relationships are free of impact for the following research

26/07/2018

2

• Raystation, Eclipse, Tomotherapy, Varian, Elekta, Velocity

Sydney NSW

API- Differences from Eclipse• WPF integration

• More developed with greater functionality

• Interactions are within the software- live updates of UI

26/07/2018

3

Promise of the future

Patient Consult

DiagnosisTherapeutic

PathwaySimulation

Contouring Planning QA

Delivery IGRTFractional Dosimetry

Deep Learning, Clustering, Nearest Neighbours

26/07/2018

4

Contouring• ABS limitations

• Similarity term is general

• No focus on regions of datasets to compare

• Increasing included datasets increases processing time

• Highly mobile tissues perform poorly

Contouring combined approach

We can then determine a “fingerprint” for individual patients as well as groups of patients

26/07/2018

5

Measurements• Using measurement metrics to drive atlas based segmentation

selection and post processing to improve final contours

196 metrics with average processing time = 70 seconds +/- 8 seconds (n=30)

26/07/2018

6

-10

-5

0

5

10

15

20

-30 -20 -10 0 10 20 30

-10

-5

0

5

10

15

-25 -20 -15 -10 -5 0 5 10 15 20 25

The Average Breast Slice (N=186)

Body Chest

26/07/2018

7

-10

-5

0

5

10

15

-25 -20 -15 -10 -5 0 5 10 15 20 25

The Average Breast Slice (N=186)

Body Chest

Contour Metrics

Closest ABS Model

Atlas Applied

New CT

3 nearest neighboursMeasure Metrics

26/07/2018

8

Contour Metrics

Closest ABS Model

Atlas Applied

New CT

3 nearest neighboursMeasure Metrics

Alternatively

Best match in entire patient database (n>200)

Post-Processing: Iterative Contour Growth

Atlas Segmentation Contract and take average HU

Expand Threshold

Remove small contoursRepeat over a given number of iterations

26/07/2018

9

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

DICE Sensitivity Precision Specificity

Breast_L

Human ABS Short ABS PP

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

DICE Sensitivity Precision Specificity

Breast_R

Human ABS Short ABS PP

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

DICE Sensitivity Precision Specificity

Liver

Human ABS Short ABS PP

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

DICE Sensitivity Precision Specificity

Heart

Human ABS Short ABS PP

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

DICE Sensitivity Precision Specificity

Lung_R

Human ABS Short ABS PP

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

DICE Sensitivity Precision Specificity

Lung_L

Human ABS Short ABS PP

Average Inter-User Variation (DICE)• Heart = 0.92• Lungs = 0.97• Breasts = 0.92

• Improvements from large atlas, to measurement directed atlas, to post processing

• PP- approaching inter-user variations DICE scores• Sensitivity > Specificity suggesting slight under-contouring

Results- Contouring Time

Incorporating a DICOM folder watch code results in auto-contouring performed before staff return from CT simulationCouch position is set from the couch height DICOM header from the CT image

26/07/2018

10

Auto-Planning• Auto-contour from CT header• Oncologist reviews and amends auto-contours• Hybrid IMRT• VMAT short arc with robust optimisation• VMAT breast/chest wall + nodes – with robust

User chooses delivery type and clinical

goals

Beam Design

Clinical Goals Added

Objective Functions

Set

Opt Parameter

Set

Run OptClinical Goals

Assessed

Increment Objective

Value

Complete Plan Presented

to User forReview

Plan Quality Check Script

Failed Goal

Counter +1

All goals pass or

Counter 10

26/07/2018

11

User chooses delivery type and clinical

goals

Beam Design

Clinical Goals Added

Objective Functions

Set

Opt Parameter

Set

Run OptClinical Goals

Assessed

Increment Objective

Value

Complete Plan Presented

to User forReview

Plan Quality Check Script

Failed Goal

Counter +1

All goals pass or

Counter 10

AutoPlan Results

*statistical significance determined as p < 0.05

0

5

10

15

20

25

Not Significant Clinical Auto

DV

H M

etr

ic

Best Performing Plan Creation

DVH Metric Comparison (n=100)

Time Measured (N=10) Min Average St Dev Max

Automated Plan (minutes :seconds) 04:48 05:31 00:42 06:21

Cl inica l Dos imetris t Time (minutes :seconds) 21:36 30:37 04:25 41:07

Relative Planning Time (%) 22% 18% 16% 15%

Time Saving (minutes) 17 25 4 35

Excluding Physician contouring of boost targets total average processing time approximately 11min

26/07/2018

12

Lt Breast Rt Breast

Auto-PlanningGood Score against recommendations

But under-coverage compared to plan database, potentially sparing lungs a little too much

26/07/2018

13

0

1000

2000

3000

0 10 20

PTV

0

200

400

0 10 20

Eval Lung

0

5000

10000

0 10 20

PTV d1

4000

4500

5000

0 10 20

PTV d98

0

50

100

0 10 20

Eval CB

4000

4500

5000

0 10 20

PTV d98

-10

-5

0

5

10

15

20

-30 -20 -10 0 10 20 30

Use of database to QA contours after auto-contouring and to predict both optimisation objectives and expected plan qualityVarious machine learning algorithms

Import

Deformation

Deformation QA

Contour Propagation

Fraction Calculation

Dose Comparison

Dose Accumulation

< 2 minute total process time

26/07/2018

14

Patient Consult

DiagnosisTherapeutic

PathwaySimulation

Contouring Planning QA

Delivery IGRTFractional Dosimetry

• The automation workflow allows for a technically feasible online adaptive process• Doctor buy in? Who reviews the deformed CTV on set?• Eventual accumulated dose statistics to feed into QUANTEC/RayBiology models and therapeutic decisions

Acknowledgments

• Mikel Byrne

• Yunfei Hu

• Trent Aland

• Yang Wang

• Emma Cai

• Nick Collett

• Guilin Liu

Ben.Archibald-Heeren@roc.team