Automated Treatment Planning Using a Database of Prior Patient Treatment Plans Todd McNutt PhD, Binbin Wu PhD, Joseph Moore PhD, Steven Petit PhD, Misha Kazhdan PhD, Russell Taylor PhD Shape DB work funded by Philips Radiation Oncology Systems McNutt Pre-Plan Point of Care Data Images Vitals Feedback in Radiotherapy Image Guidance Tracking/ Gating Patient Information Breathing Motion Marker Tracking Real Time In Treatment Knowledge Guidance Adaptive RT TxCT Orthogonal X-ray Multi-TxCT MRI In Patient In Disease Clinical Experience Dosimetry Outcome Toxicity Clinical Trials Closed Loop Medicine Population Information Oncospace: An escience program for the advancement of care in radiation oncology Objectives: To develop an analytical database and infrastructure to store clinical information for personalized medicine and future analysis Project : Integration of Data Collection with Clinical Workflow Project : Database Design: Security and Distributed Web-Access Project 3: Tools for Query, Analysis, Navigation and Decision Support Project : Data Mining, Decision Support and Biostatistic Research, McNutt
Tumor information is stored including staging and image basedrecistassessment. Personal Health Information is stored in a single table to facilitateanonymization. Toxicity andoutcomes. Labvalues Chemotherapy, medications, andsurgeries Patient History Regions of interest are stored as run-length encoded masks associated with each Patient Representation. Shape and shape relationship descriptors are stored for fast query of patient shapesimilarities. DoseVolumeHistograms arestoredfor eachregionofinterestfor fast query. DVH stored for both treatment summary and individual treatmentsessions Dose distributions are stored for each radiotherapy session. Each radiotherapy session is associated with a single patient representation. The transformation table stores the transformationbetweenmultiplepatientrepresentations enablingdoseaccumulation. Oncospace Website Grade & 3 Influence of Shape Shape Characteristics Volume Positional relationships between structures Separation of surfaces Shape Change in Time Influence Plan quality (IMRT) Ability to achieve Tx goal Motion management Toxicity Simplification of information without loss of relevance? Patient Shape Clusters
Use of ROI Shapes in Oncospace Overlap Volume Histogram Shape Relationship Descriptor DB of prior patients far 8 norm overlap volume L: OVH58.9 L:OVH63 L:OVH7.8 R:OVH58 R:OVH63.7 R:OVH7.6.5..3.. - - 3 5 6 7 8 99.5 Expansion Distance (cm) Decisions: Plan Quality Assess Automated IMRT Expected Toxicities Dosimetric Trade-offs OVH for similar patients to predict expected DVH DVH predict toxicity? normalized volume.9.8.7.6.5..3.. d3:ovh of PTVH 6-8 Right parotid Left parotid 5 5 5 3 35 5 5 55 6 65 7 6 d:ovh of PTVM DVH Prediction close - - - d:ovh of PTVL 3 Dose (Gy) McNutt 9 Overlap That Volume was descriptive Histogram only (OVH) OVH maps the shape of an OAR to a volume-distance plane Actual through computation target expanding is with a Euclidean and contracting Distance Transform Organ Algorithm Target which is more efficient than amm the process described. Volume amm amm Michael Kazhdan, Patricio Simari, Todd McNutt, Binbin Wu, Robert Jacques, Ming Chuang, and Russell Taylor, A Shape Relationship Descriptor for RadiationTherapy Planning Medical Image Computing and Computer-Assisted Intervention 576/9(), 8 (9) amm -a mm -a mm a mm a mm Distance 3
From a point to an organ Which dose? Maximum dose? Dose to e.g. 5% DVH OAR PTV OAR 5% volume at.5 cm from PTV Overlap Volume Histogram describes distance from a (sub)volume to the PTV An example of OVH OAR OAR For parallel organs, OAR (red) is more easily spared. For serial organs, OAR (blue) is more easily spared. Comparison between L and 5L: L is an outlier Patient Patient 5 3-D shapes of the left parotid and PTV58.. 3-D shapes of the left parotid and PTV63..9.8.7.6.5..3...9.8.7.6.5..3.. L:OVH58 L:OVH63 L:OVH7 5L:OVH58 5L:OVH63 5L:OVH7-3 5 6 7 8 Distance (cm) OVH of L and 5L L:re-plan L:original 5L:original 3-D shapes of the left parotid and PTV7. 3 5 6 7 8 Dose(Gy) DVH of L and 5L
Re-plan results of patient Original plan Re-plan.8.6.5.. Treatment Plan Quality Control (outlier detection): parotids OVH L OVH M OVH H.8.6.5.. DVH of parotid - Distance (cm) Clinical goal: Parotid: V(3Gy)<5% of volume D V H = T [ O V H, O V H, O V H ] 6 8 Distance (cm) Dose corresponding 5% of volume: 6 8 Distance (cm) L M H Distance corresponding 5% of volume: DGy = T [ d Lcm, d M cm, d H cm] 6 8 Dose (Gy) Detection rule: For covering the same percentage volume of the OAR, the larger the expanded distance is, the easier to spare the OAR. Treatment Plan Quality Control (outlier detection): parotids 8 d3:ovh of PTVH 6-8 6 d:ovh of PTVM - - - d:ovh of PTVL Dose corresponding 5% of volume: : 5Gy : 5 3Gy : 3 35Gy : 35 Gy + : 5Gy :5 5Gy :5 55Gy :55 6Gy : >6Gy 3 5
Re-plan results of the 7 outlier patients V(3Gy).7.65.6.55.5.5..35 original: V(3Gy) re-plan: V(3Gy) Clinical goal (parotid): V(3Gy)<.5 V(3Gy).9.8.7.6.5. original: V(3Gy) re-plan: V(3Gy).3 3 5 6 7 8 9 3 5 6 7.3 3 5 6 7 8 9 7 outlier parotids 9 non-outlier parotids 6 re-plan patients, 7 are outliers, 9 non-outlier indicated sby the OVH. V(3Gy) of 8 parotids among the 7 outlier parotids are reduced to below 5%! Non-outliers were not significantly improved All 6 re-plans are reviewed by physician. Predict dose minimal dose to e.g. 5% of organ Lookup distance PTV to 5% of the organ Read OVHs of all prior patients Select all patients for which the 5% was closer to the PTV Lookup dose at 5% DVH for selected patients Select the lowest achieved dose = prediction of what is achievable New pt More difficult Less difficult OVH-based Prediction 3 5 brain.. L parotid cord+ Achievable DVHs for a new patient IMRT Optimization 3 brainstem mandible Comparing OVH Predicting DVH IMRT optimization 6
Use of SQL DB reduces search to an SQL query H&N Retrospective Planning Demonstration 5 random pts from a DB of 9 H&N pts for OVHassisted planning demonstration IMRT-SIB: 58. Gy, 63 Gy and 7 Gy DVH objectives of 3 OARs queried from the DB as initial planning goals in a leave-one-out manner Dosimetry of 3 sets of plans were compared: CP - Clinical plans OP - OVH-assisted plans after optimization OP - Final OVH-assisted plans 5 pts: PTV comparisons among CP,OP and OP PTV coverage and homogeneity were slightly better in both OPs; conformity was similar. 7
5 pts: OAR Sparing among CP,OP and OP Significantly lower in both OPs: cordmm (~6 Gy), brainstem (~7. Gy) and contra-lateral parotid (~7%) Re-plan results of patient Red: re-plan Blue: original Dot: right parotid No-dot: left parotid.9.8.7.6.5..3.. 3 5 6 7 8 Dose(Gy) Re-plan results for other OARs Patient brain (Gy) Brainstem (Gy) Cordmm (Gy) L inner ear (Gy) original 6.5 5.58.75 57.8 re-plan 56.33 6.8 37.89 3.7 Patient R inner ear (Gy) mandible (Gy) larynx for edema esophagus (Gy) original.57 66.58 6% 63.7 Re-plan 38.38 63.78 59% 6 Brain, brainstem, cordmm, esophagus and mandible: maximal dose Inner ear: mean dose Larynx for edema: V(5Gy) Plan comparison: efficiency (5 plans) Average number of optimization rounds per OP is.9; that number for a CP is 7.6; 3 OPs finished in a single round. 8
Prospective clinical trial study Binbin Wu PhD, Giuseppe Sanguineti MD - IRB Approved Purpose: Explore the feasibility of the automated OVH planning tool in clinic. Pt accrual: Pts accrued from 7/ / (6 oropharynx; 9 larynx; 5 nasopharynx) Protocol: Definitive IMRT to 7 Gy in 35 Fractions to GTV and 63 Gy and 58. Gy to high and low risk CTVs. Three PTVs for each pt: PTV58., PTV63 and PTV7 Volume in cc 5 3 Volume distribution of pts PTV 58. PTV 63 PTV 7 9 8 7 6 5 3 3 5 7 9 3 5 7 9 3 5 7 9 3 33 35 37 39 Patient No. PTV 7 : mean (73.cc); median (3.78 cc); SD (53 cc). PTV 63 : mean (363.5 cc); median (39. cc); SD (.6 cc). PTV 58. : mean (9. cc); median (9.7cc); SD (53.7cc) Study work flow plans generated for each Patient New patient: contours of the OARs and CTVs CP- Clinical Plan manually created by dosimetrists (unaware of the study). AP- Automated Plan plans are automatically generated by the proposed TPS. Clinical planning is guided by Dr. Sanguineti and in-house dosimetric guidelines. week of post-approval of CP, both AP and CP are blindly reviewed by Dr. Sanguineti. One of the plans is chosen as the better one. 9
Dosimetric Results: CP vs. AP Primary OARs (optic nerve, chiasm, brainstem, brain, cord and mandible) AP: reduced by. Gy (p=.) overall PTV coverage (V 95 in %) AP: increased by.6% (p=.) overall Secondary OARs (parotid, brachial plexus, larynx, inner ear, oral mucosa, esophagus AP: reduced by.6 Gy (p=.) overall PTV homogeneity and conformity AP: significant better homogeneity in PTV 63 (p=.) and PTV 7 (p <.) AP: significant better conformity in PTV 58. (p=.9). AP: fully automated plans CP: clinical plans manually created by dosimetrists in their regular way AP: optimization runs per plan (~3 minutes) CP: ~ (SD: 9) optimization runs per plan No. of clicks in CPs 3 Planning efficiency 9 8 7 6 5 3 3 5 7 9 3 5 7 9 3 5 7 9 3 33 35 37 39 Patient No. oropharynx larynx nasopharynx Physician Preference Dr. Sanguineti reviewed the isodose distributions and DVH curves without knowing the origins of the plans. Based on his opinion, All APs (/) are clinically acceptable and can be used to treat patients 7/ APs are clinically superior to the CPs
73.5 Gy 63 Gy 7 Gy 58. Gy 5 Gy AP completed in. minutes Both clinically acceptable; physician preferred CP due to less hot spots inside PTV7 although better organ sparing in AP Dash curve: AP Solid curve: CP Pancreas example Steven Petit Mean kidney dose decrease 6 Gy! Results: liver x Original plan predicted Replanned predicted 5% constraint: 9% within Gy 96% within Gy 5% constraint: 86% within Gy % within Gy 65% constraint: 93% within Gy % within Gy After replanning: decrease in mean dose = % [ 7%]
Results: Kidneys 5% constraint: 89% within Gy % within Gy 5% constraint: 96% within Gy 96% within Gy 75% constraint: 65% within Gy 87% within Gy After replanning: decrease in mean dose = 7% [ 76%] Does this really work? One modification Consider part of organ within beams-eye-view of the beams (+ margin) head Liver PTV Field borders (Beams-eye-view + margin) toes Clinical Release Joseph Moore Tool is easily adapted to any site Release for Pancreas first Standardization of ROI Names Standardization of technique to some extent Query DB for predicted dose level for each objective function Completed new plans push to DB ROI Name Mapping IMRT Objective Function Query Green boxes queried from DB Red if failed Developed by Joe Moore
Autoplanned Joseph Moore + Dosimetry Summary Automated TPS without user intervention OVH: retrieve geometrically similar pts DB of prior plans: control plan quality of future plans Quality of new plans is independent of experience of planners; consistent with quality of prior plans in DB Clinical trade-offs made by physician are captured in the database Easily implemented to other disease sites (pancreas and prostate) Easily implemented to VMAT modality (used current DB for VMAT) Easily applied with any commercial TPS JHU - CS Russ Taylor PhD Misha Kazhdan PhD Patricio Simari PhD JHU - Physics Alex Szalay PhD Philips PROS Karl Bzdusek Erasmus/MAASTRO Steven Petit PhD Acknowledgments JHU-RO Binbin Wu PhD Kim Evans MS Joseph Moore PhD Wuyang Yang MS, MD Giuseppe Sanguinetti MD Russell Hales MD Joseph Herman MD John Wong PhD Theodore DeWeese MD 3