Creating a Knowledge Based Model using RapidPlan TM : The Henry Ford Experience

Transcription

1 DVH Estimates Creating a Knowledge Based Model using RapidPlan TM : The Henry Ford Experience Karen Chin Snyder, MS, DABR AAMD Region V Meeting October 4, 2014

2 Disclosures The Department of Radiation Oncology at Henry Ford Hospital has a research agreement with Varian Medical Systems. The speaker has received speakers honoraria from Varian Medical Systems

3 Varian Medical Systems Fair Balance Safety Statement INTENDED USE STATEMENT Varian Medical Systems linear accelerators are intended to provide stereotactic radiosurgery and precision radiotherapy for lesions, tumors, and conditions anywhere in the body where radiation treatment is indicated. SAFETY Radiation treatments may cause side effects that can vary depending on the part of the body being treated. The most frequent ones are typically temporary and may include, but are not limited to, irritation to the respiratory, digestive, urinary or reproductive systems, fatigue, nausea, skin irritation, and hair loss. In some patients, they can be severe. Treatment sessions may vary in complexity and time. Radiation treatment is not appropriate for all cancers. Medical Advice Disclaimer Varian as a medical device manufacturer cannot and does not recommend specific treatment approaches. Individual treatment results may vary. For more information, visit MyVarian.com. FOR HEALTH CARE PROFESSIONALS ONLY

4 Why use Knowledge Based Models? Inverse Planning Variability Dependent on Planner s skill Balancing clinical objectives Iterative process that can be time consuming

5 RapidPlan TM Varian s Solution for Knowledge-Based Planning DVH Estimation Model Tumor Site Specific Trained from existing plans that have accepted OAR tradeoffs Uses principal component analysis for each OAR Different sub-volumes use different DVH estimation Automatic objective generation DVH estimates Automatic objectives The estimated DVH PC coefficient From the structure set of current plan

6 Basic Steps in Creating a KB Model Add Plans to the Model Train and Verify the Model Validate the Model

7 Basic Steps in Creating a KB Model Training Data Set Add Plans to the Model Train and Verify the Model Validation Data Set Validate the Model

8 Adding Plans to the Model Add Plans to the Model Geometry of Targets & OARs Plan Quality Treatment Technique Dose Prescription

9 Adding Plans to the Model Add Plans to the Model Geometry of Targets & OARs Plan Quality Treatment Technique Dose Prescription

10 Geometry of Targets and OARs - Size, Contouring and Physiology of OARs How do you treat at your clinic? Example: Rectum empty, bladder full and PTV expansions - change the overlap region between the OARs and Target - range of volumes of OARs in the model ie, full rectum as outlier

11 Geometry of Targets and OARs - Position of OARs relative to target Amount of Overlap Example: HN: amount of overlap with parotids Lung: RTOG 0915, max doses less than Rx dose. No overlap. - Differing volumes in the in-field region, all scenarios

12 Geometry of Targets and OARs - Position of OARs relative to target Distance of Target to OARs - Differing out-of-field volumes OAR s proximity to Spinal Cord

13 Geometry of Targets and OARs - Size, Contouring and Physiology of OARs Which OARs do you want the model to support? Example: Stomach and Brachial Plexus - Typical location of target compared to extreme locations - Need sufficient number of cases to train the model to give an estimate - OARs can be included without an estimation (fixed objective)

14 Adding Plans to the Model Add Plans to the Model Geometry of Targets & OARs Plan Quality Treatment Technique Dose Prescription

15 Treatment Technique - VMAT, IMRT, both - Gantry Angles - Non-coplanar, coplanar Beam geometry Important in Structure Volume Partitioning - Different out-of-field/in-field/ leaf-transmission and target overlap volumes Target overlap region In-field region Transversal view Leaf-transmission region Out-of-field region Image courtesy of Esa Kuusela, Varian Medical Systemes

16 Adding Plans to the Model Add Plans to the Model Geometry of Targets & OARs Plan Quality Treatment Technique Dose Prescription

17 Dose Prescription - Simultaneous Boost (SIB) - Attach multiple PTVs to the structure - Differing Dose Prescriptions - Scaling of doses - Not ideal for different fractionation schemes S.H. Benedict, K.M. Yenice, D. Followill, et al. Stereotactic body radiation therapy: The report of AAPM Task Group 101 Med Phys, 37 (2010), pp

18 Adding Plans to the Model Add Plans to the Model Geometry of Targets & OARs Plan Quality Treatment Technique Dose Prescription

19 Plan Quality Physician Preferred Dose Distributions PTV coverage - NTO used, Shoulder of DVH, allowable hot spot OAR Sparing Example: Rib sparing in HFHS Lung SBRT Model

20 How Many Plans to Include in the Model Complexity of model, and what you want it to include - Prostate ~40, Prostate w/ LN ~ 70, HN SIB ~ 116, SBRT Lung = 105 Overlap of Data points - Basically the same plan/location - Can always add new interesting cases to the model, will need to re-validate

21 Basic Steps in Creating a KB Model Add Plans to the Model Train and Verify the Model Validate the Model

22 Train and Verify the Model Train the Model Re-Extract Identify Outliers Add Similar Plans Re-plan Re-contour Dosimetric Geometric Remove Plan Evaluate if Plan is Appropriate Neither?

23 DVH Tool Validation Tools - OAR of each training patient and corresponding estimate Estimate

24 Training Log - Outlier Summary and Metrics - Statistics Summary Validation Tools

25 Validation Tools Regression Plot - DVH PC1 vs Geometric Distribution PC1 2 Std Devs Model Fit

26 Training log Identifying Geometric Outliers - Cook s Distance (CD) : indicates influential point, alerts you when >4 - Modified Z score (mz) 4 : difference between individual and mean Regression Plot: DVH PC1 vs Geometric Distribution PC1 Contouring issues: partially contoured structure

27 Training log Identifying Dosimetric Outliers - Studentized Residual (SR) : measures difference between original and estimated data, threshold of >3 will alert - Areal difference of estimate (da) difference in area between estimated and actual DVH Residual Plot: DVH PC vs Estimated DVH PC1 Original Plan is Higher than Estimate Example: Atypical tradeoff between OARs Original Plan is lower than Estimate

28 Training log Neither? - Cook s Distance (CD) : indicates influential point - Studentized Residual (SR) : measures difference between original and estimated data Regression Plot: DVH PC1 vs Geometric Distribution PC1 Need more data points to fill in regression plot

29 Goodness- of-fit Statistics Overfitting - Chi square: range 0-1, want to be closer to 1 - Coefficient of determination: close to 1

30 Basic Steps in Creating a KB Model Add Plans to the Model Train and Verify the Model Validate the Model

31 Validate the Model The Model Patient Plan Model Estimate Final Dose Calc Estimated Plan Validate DVH Estimation Validate Automatic Priorities Validate the Model

32 Patient Plan vs Estimate For a particular patient geometry, is your model able to give you a good estimate? Validate the DVH Estimation - Do you need to go back and evaluate your training set/re-train model?

33 Model Estimate vs Estimated Plan After optimization and final calculation Validate the Objectives - Are the estimated priorities able to achieve the estimated DVH?

34 Patient Plan vs Estimated Plan How well is your model performing? Are you obtaining clinically acceptable plans? Validate the Overall Model - Use a good range of geometries - Are there exceptions/limitations after validation?

35 Limitations in our Model Issue in cases where the PTV abuts the heart Max heart dose is 34Gy, 95% of PTV to receive 48Gy Needs additional user modification in order to meet the OAR max dose constraint Plan may not meet other planning criteria, GI, 99% PTV coverage, max dose from PTV+2cm Not supported in the current model Fixed PTV objectives Not many patient cases

36 Published Models - Varian Provided Models Publication and Sharing - Intended usage and applicability to your clinic - You can modify Varian Models - Need to be independently validated by you before application to clinic - Easily share models

37 Using a RapidPlan Model Integrated part of the Photon Optimizer - Choose the Model - Match OARs to Model Structure - Match PTV to PTV, and set Target Dose - Generate Estimates and Objectives and optimize. Similar to applying a fixed objective template, but with the knowledge of patient geometry and experience of previous planners

38 Knowledge based models Conclusion - Share knowledge with DVH estimation and automatic objective creation - 3 basic steps in creating your own model - You can create your own models or modify existing models How KB Models can help you - Quality Control of IMRT plans - Standardization - Automation - Inter-institutional sharing

39 Thank You The KB Team at HF: Indrin Chetty, PhD Ajlouni Munther, MD Salim Siddiqui, MD, PhD Jinkoo Kim, PhD Anne Reding, CMD Corey Fraser, CMD Varian: Helen Phillips, Esa Kuusela