Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Location Prediction of Moving Target Dan Ruan Department of Radiation Oncology Stanford University AAPM 2009 Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Outline of Topics 1 Motivation and Problem Definition Motivation Problem Description Two Perspectives Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Outline of Topics 1 Motivation and Problem Definition Motivation Problem Description Two Perspectives 2 Deterministic Perspective Sample Path Estimation via Motion Modeling Deterministic regression Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Outline of Topics 1 Motivation and Problem Definition Motivation Problem Description Two Perspectives 2 Deterministic Perspective Sample Path Estimation via Motion Modeling Deterministic regression 3 Stochastic Perspective Estimation of Stochastic Process Stochastic Regression: Recent Development Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Outline of Topics 1 Motivation and Problem Definition Motivation Problem Description Two Perspectives 2 Deterministic Perspective Sample Path Estimation via Motion Modeling Deterministic regression 3 Stochastic Perspective Estimation of Stochastic Process Stochastic Regression: Recent Development 4 Summary Disclaimer: unless indicated otherwise, the views in this talk represent personal opinion, not an official statement of AAPM or Stanford University. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Motivation Problem Description Two Perspectives Outline 1 Motivation and Problem Definition Motivation Problem Description Two Perspectives 2 Deterministic Perspective Sample Path Estimation via Motion Modeling Deterministic regression 3 Stochastic Perspective Estimation of Stochastic Process Stochastic Regression: Recent Development 4 Summary Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Motivation Problem Description Two Perspectives Motivation • Adaptive therapy with online motion compensation At each time instant, requires Target localization ⇒ Treatment adaptation. • System latency is ubiquitous: necessitates prediction (estimation with delayed observations). [Motion Monitoring + Prediction] ⇒ Treatment adaptation • From a filtering perspective, prediction (extrapolation) is a viable option for handling asynchronous multiple observation streams in hybrid systems. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Motivation Problem Description Two Perspectives Problem description Observation Event I Observation Event II Processing I Observation Event III Processing III Processing II Treatment II Treatment I Treatment III Time System Latency Observation Interval Goal: given an observed sample path {si }K i=1 , to causally estimate the underlying state sK +L at a future time (times). Talk will focus on intra-fraction respiratory motion, idea generalizes. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Motivation Problem Description Two Perspectives Key factors observation (sample path) mode - usually discrete {si }. ◦ sampling pattern (uniform v.s. non-, rate) ◦ uncertainty ◦ missing samples/outliers? detectable? prediction requirement: ◦ setup: prediction horizon pattern (fixed length, varying, continuous) ◦ accuracy requirement state definition: ◦ instantaneous target dynamics (location, velocity, etc., with possibly categorical classification). ◦ instantaneous anatomical configuration. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Motivation Problem Description Two Perspectives Two perspectives • Deterministic perspective Consider the underlying state as deterministic unknown ⇒ estimate its value , explicitly seek ŝ. ◦ sample path modeling: most intuitive. ◦ deterministic regression: estimates a deterministic input/output map from the observed portion of sample path. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Motivation Problem Description Two Perspectives Two perspectives • Deterministic perspective Consider the underlying state as deterministic unknown ⇒ estimate its value , explicitly seek ŝ. ◦ sample path modeling: most intuitive. ◦ deterministic regression: estimates a deterministic input/output map from the observed portion of sample path. • Stochastic perspective Consider the underlying state as stochastic unknown ⇒ estimate its distribution. ◦ random process modeling: most stochastic filtering schemes ◦ ... Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Motivation Problem Description Two Perspectives Categorization of Prediction Schemes Sample Path Est. Deterministic Stochastic Dan Ruan Prediction Regression Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Motivation Problem Description Two Perspectives Categorization of Prediction Schemes Sample Path Est. Deterministic Stochastic Cosine, Fourier, Wavelet, Piecewise Linear, ARMA Stochastic filter, KF, PF Dan Ruan e.g., Prediction Regression ANN, SVM, LOESS ... Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Outline 1 Motivation and Problem Definition Motivation Problem Description Two Perspectives 2 Deterministic Perspective Sample Path Estimation via Motion Modeling Deterministic regression 3 Stochastic Perspective Estimation of Stochastic Process Stochastic Regression: Recent Development 4 Summary Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Sample path estimation via motion modeling • Sample path modeling: generally seeks a complete temporal-spatial description of the underlying evolving state by modeling the feasible sample path. Any method that reminds you of a sketching procedure... • Major playground: representation e.g. modified cosine model[Lujan 94], local linear model[Wu 04], wavelet + AR[Ernst 07], real-time profiling[Ruan 09] Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Local linear representation • A normal breathing cycle is decomposed into three stages: inhale(IN), exhale(EX), and end of exhale (EOE). • Periodic progression in/out of each linear segment is modeled with a discrete set of rules, local parameters estimated. • Potential local continuous prediction capability. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Profiling representation • The observations are considered as noisy discrete samples from a periodic shape subject to phase warping and baseline drift. si = [fT (φ(t)) + b(t) + w (t)]|t=ti . • The online profiling system estimates the instantaneous shape fT , phase warping φ (thus frequency variation) and baseline drift b in real-time. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Profiling schematic • robust to nonuniform samples, high observation noise, outliers. • provides functional descriptor with natural continuous horizon prediction. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Profiling applied to prediction (cont.) 10 8 6 5 displacement displacement 4 0 −5 2 0 −2 −4 −6 −10 −8 −15 0 5 10 15 Time (sec) 20 25 −10 0 30 5 10 15 Time (sec) 20 25 30 5 10 15 Time (sec) 20 25 30 20 15 15 10 displacement displacement 10 5 0 5 0 −5 −5 −10 −10 0 −15 5 10 15 Time (sec) 20 25 30 0 available observation, unknown observation, prediction with profiling extrapolation. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Profiling applied to prediction 6 2 4 −2 displacement displacement 0 −4 −6 −8 0 −2 −4 −10 −12 0 2 −6 5 10 15 Time (sec) 20 25 0 30 5 10 15 Time (sec) 20 25 30 15 Time (sec) 20 25 30 15 5 10 0 displacement displacement 5 0 −5 −10 −5 −10 −15 −15 −20 −20 0 5 10 15 Time (sec) 20 25 30 −25 0 5 10 available observation, unknown observation, prediction with profiling extrapolation. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Deterministic regression for state prediction • Location prediction with deterministic regression: assumes a consistent relationship between a recent observed temporal segment and the state after a fixed latency. The goal is to estimate such a consistent map. Any method that deterministically “learns” from input-output pairs. • Major playground: functional approximation and estimation. e.g. LOESS [Ruan 06], ANN[Murphy 06], SVM[Ernst 08], ANFIS [Kakar, 05] Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Generating input-output pairs for regression Lookahead Length Response Unavailable Current SK Delay SK−∆ 250 1 200 0.8 150 0.6 100 0.4 50 0 0 0.2 2 4 6 8 10 Time (seconds) 12 14 16 Regression Weights Distance in State Space Current State xK Actual yK (unavailable) 0 18 Dan Ruan • input xi = [si−(p−1)∆ , . . . , si−∆ , si ], concatenation of p delayed samples • output yi = si+L , discrete lag L corresponds to prediction length • With most current sample sK , training set {(xi , yi )}i<K −L . Testing input: xK , want ŷK . • Goal: learn map g : <p → <, y ≈ g (x) from training set, especially around input xK . Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Functional estimation Parametric: LOESSP • assume g (x) = Q q=1 βq zq (x) with tensor basis zq . • determine local inference weight Nonparametric: ANN, ANFIS −1 wi = κ(hK kxi − xK k). • estimate the coefficient β via WLS. β̂ = arg min β K −L X i=1 Q X wi (yi − βq zq (xi ))2 . q=1 Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Sample Path Estimation via Motion Modeling Deterministic regression Performance of deterministic regression 600ms lookahead length, RPM relative displacement. • ANN: more flexible, favorable with abundant training samples. • LOESS: more specific functional form, more robust with sparse samples, computationally efficient. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Estimation of Stochastic Process Stochastic Regression: Recent Development Outline 1 Motivation and Problem Definition Motivation Problem Description Two Perspectives 2 Deterministic Perspective Sample Path Estimation via Motion Modeling Deterministic regression 3 Stochastic Perspective Estimation of Stochastic Process Stochastic Regression: Recent Development 4 Summary Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Estimation of Stochastic Process Stochastic Regression: Recent Development Starting from KF • iterates between predicting and updating the R.V. dist. • the performance limitation of KF in respiratory prediction lies in linearity, NOT its stochastic nature. • motivates stochastic estimation with more general dynamic model promising. Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Estimation of Stochastic Process Stochastic Regression: Recent Development What about stochastic regression A recent development • admits the random state, relaxes requirement for map consistency. • nonparametrically learns distribution. 4.5 4 3.5 RMSE 3 2.5 2 MRS Linear 1.5 1 ANN Kernel 0.5 0 5 10 15 20 sampling frequency (Hz) Dan Ruan Prediction 25 30 Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Outline 1 Motivation and Problem Definition Motivation Problem Description Two Perspectives 2 Deterministic Perspective Sample Path Estimation via Motion Modeling Deterministic regression 3 Stochastic Perspective Estimation of Stochastic Process Stochastic Regression: Recent Development 4 Summary Dan Ruan Prediction Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary deterministic samp. path regression utilize periodicity nonunif. sample local variation cont. prediction accuracy needs work Y needs work Y Y Dan Ruan Y N N N Y Prediction stochastic process regression needs work Y Y Y needs work Y N Y N Y Outline Motivation and Problem Definition Deterministic Perspective Stochastic Perspective Summary Summary • Motion prediction is a challenging problem - a lot of room for improvement. • Different interpretations/perspectives offer new opportunities. • Proper choice depends on specific applicationl • Play with different modifications, combinations and have fun! Thank you! Dan Ruan Prediction