Optimal Design with Automatic Differentiation: Exploring Unbiased Steady-State Relaxometry Jan 13, 2014 Jason Su.

Slides:



Advertisements
Similar presentations
Yi Heng Second Order Differentiation Bommerholz – Summer School 2006.
Advertisements

Complex Data For single channel data [1] – Real and imag. images are normally distributed with N(S, σ 2 ) – Then magnitude images follow Rice dist., which.
Running a model's adjoint to obtain derivatives, while more efficient and accurate than other methods, such as the finite difference method, is a computationally.
Design of Experiments Lecture I
July 31, 2013 Jason Su. Background and Tools Cramér-Rao Lower Bound (CRLB) Automatic Differentiation (AD) Applications in Parameter Mapping Evaluating.

Journal Club: mcDESPOT with B0 & B1 Inhomogeneity.
FTP Biostatistics II Model parameter estimations: Confronting models with measurements.
Variance reduction techniques. 2 Introduction Simulation models should be coded such that they are efficient. Efficiency in terms of programming ensures.
More MR Fingerprinting
Relaxometry & Image Processing Technical Update Clinical Findings/Product Need Competitive Info Recommendations T1 (DESPOT1 and LL) parameter fitting via.
Gizem ALAGÖZ. Simulation optimization has received considerable attention from both simulation researchers and practitioners. Both continuous and discrete.
Classification and Prediction: Regression Via Gradient Descent Optimization Bamshad Mobasher DePaul University.
Infinite Horizon Problems
Validation and Monitoring Measures of Accuracy Combining Forecasts Managing the Forecasting Process Monitoring & Control.
Active Calibration of Cameras: Theory and Implementation Anup Basu Sung Huh CPSC 643 Individual Presentation II March 4 th,
458 Interlude (Optimization and other Numerical Methods) Fish 458, Lecture 8.
Motion Analysis (contd.) Slides are from RPI Registration Class.
L Berkley Davis Copyright 2009 MER301: Engineering Reliability Lecture 14 1 MER301: Engineering Reliability LECTURE 14: Chapter 7: Design of Engineering.
Copyright © 1998 Wanda Kunkle Computer Organization 1 Chapter 2.1 Introduction.
Theoretical & Industrial Design of Aerofoils P M V Subbarao Professor Mechanical Engineering Department An Objective Invention ……
CISE-301: Numerical Methods Topic 1: Introduction to Numerical Methods and Taylor Series Lectures 1-4: KFUPM.
Classification and Prediction: Regression Analysis
Ch 8.1 Numerical Methods: The Euler or Tangent Line Method
Efficient Model Selection for Support Vector Machines
Computation of the Cramér-Rao Lower Bound for virtually any pulse sequence An open-source framework in Python: sujason.web.stanford.edu/quantitative/ SPGR.
CISE-301: Numerical Methods Topic 1: Introduction to Numerical Methods and Taylor Series Lectures 1-4: KFUPM CISE301_Topic1.
Analytical vs. Numerical Minimization Each experimental data point, l, has an error, ε l, associated with it ‣ Difference between the experimentally measured.
Improved Gene Expression Programming to Solve the Inverse Problem for Ordinary Differential Equations Kangshun Li Professor, Ph.D Professor, Ph.D College.
Numerical Methods Applications of Loops: The power of MATLAB Mathematics + Coding 1.
1 Performance Analysis of Coexisting Secondary Users in Heterogeneous Cognitive Radio Network Xiaohua Li Dept. of Electrical & Computer Engineering State.
Approximation Algorithms Pages ADVANCED TOPICS IN COMPLEXITY THEORY.
An Introduction to Programming and Algorithms. Course Objectives A basic understanding of engineering problem solving process. A basic understanding of.
CRLB via Automatic Differentiation: DESPOT2 Jul 12, 2013 Jason Su.
EE369C Final Project: Accelerated Flip Angle Sequences Jan 9, 2012 Jason Su.
My talk describes how the detailed error diagnosis and the automatic solution procedure of problem solving environment T-algebra can be used for automatic.
Research Update: Optimal 3TI MPRAGE T1 Mapping and Differentiation of Bloch Simulations Aug 4, 2014 Jason Su.
ECE 8443 – Pattern Recognition ECE 8423 – Adaptive Signal Processing Objectives: Deterministic vs. Random Maximum A Posteriori Maximum Likelihood Minimum.
Optimal Experimental Design Theory. Motivation To better understand the existing theory and learn from tools that exist out there in other fields To further.
Optimal SSFP Pulse-Sequence Design for Tissue Density Estimation Zhuo Zheng Advanced Optimization Lab McMaster University Joint Work with C. Anand, R.
Disclosure risk when responding to queries with deterministic guarantees Krish Muralidhar University of Kentucky Rathindra Sarathy Oklahoma State University.
MA/CS 375 Fall MA/CS 375 Fall 2002 Lecture 31.
CP Summer School Modelling for Constraint Programming Barbara Smith 2. Implied Constraints, Optimization, Dominance Rules.
1 Markov Decision Processes Infinite Horizon Problems Alan Fern * * Based in part on slides by Craig Boutilier and Daniel Weld.
Algorithms & Flowchart
September Bound Computation for Adaptive Systems V&V Giampiero Campa September 2008 West Virginia University.
J OURNAL C LUB : Lankford and Does. On the Inherent Precision of mcDESPOT. Jul 23, 2012 Jason Su.
Solution of a Partial Differential Equations using the Method of Lines
Lecture 4: Statistics Review II Date: 9/5/02  Hypothesis tests: power  Estimation: likelihood, moment estimation, least square  Statistical properties.
1 Markov Decision Processes Infinite Horizon Problems Alan Fern * * Based in part on slides by Craig Boutilier and Daniel Weld.
Copyright © The McGraw-Hill Companies, Inc. Permission required for reproduction or display. 1 Chapter 3.
J OURNAL C LUB : Deoni et al. One Component? Two Components? Three? The Effect of Including a Nonexchanging ‘‘Free’’ Water Component in mcDESPOT. Jan 14,
559 Fish 559; Lecture 5 Non-linear Minimization. 559 Introduction Non-linear minimization (or optimization) is the numerical technique that is used by.
1 1 Slide Simulation Professor Ahmadi. 2 2 Slide Simulation Chapter Outline n Computer Simulation n Simulation Modeling n Random Variables and Pseudo-Random.
J OURNAL C LUB : Lin and Song. (Philips and UPenn) Improved Signal Spoiling in Fast Radial Gradient-Echo Imaging: Applied to Accurate T1 Mapping and Flip.
Ch 8.2: Improvements on the Euler Method Consider the initial value problem y' = f (t, y), y(t 0 ) = y 0, with solution  (t). For many problems, Euler’s.
“Joint Optimization of Cascaded Classifiers for Computer Aided Detection” by M.Dundar and J.Bi Andrey Kolobov Brandon Lucia.
Onlinedeeneislam.blogspot.com1 Design and Analysis of Algorithms Slide # 1 Download From
Anders Nielsen Technical University of Denmark, DTU-Aqua Mark Maunder Inter-American Tropical Tuna Commission An Introduction.
Neural Networks The Elements of Statistical Learning, Chapter 12 Presented by Nick Rizzolo.
A CCELERATED V ARIABLE F LIP A NGLE T 1 M APPING VIA V IEW S HARING OF P SEUDO -R ANDOM S AMPLED H IGHER O RDER K-S PACE J.Su 1, M.Saranathan 1, and B.K.Rutt.
Introduction to Machine Learning, its potential usage in network area,
Bounded Nonlinear Optimization to Fit a Model of Acoustic Foams
Fisher Information Matrix of DESPOT
Examining mcDESPOT Mar 12, 2013 Jason Su.
Non-linear Minimization
Recent Advances in Iterative Parameter Estimation
Spectral methods for stiff problems MATH 6646
BSSFP Simulations The goal of these simulations is to show the effect of noise on different pcSSFP reconstruction schemes Further examine the uniformity.
CISE-301: Numerical Methods Topic 1: Introduction to Numerical Methods and Taylor Series Lectures 1-4: KFUPM CISE301_Topic1.
Presentation transcript:

Optimal Design with Automatic Differentiation: Exploring Unbiased Steady-State Relaxometry Jan 13, 2014 Jason Su

Outline Concept and theory Practical usage with automatic differentiation (AD) Exploring relaxometry with SPGR and bSSFP

Motivation I started with the goal of trying to optimize mcDESPOT and the estimate of MWF – Lankford et al. criticized the theoretical accuracy of mcDESPOT based on CRLB analysis – Interestingly however, this criticism inspired the approach to finding the solution Since then, I’ve discovered that this project can have greater possibilities – The following provides a framework that is suitable for the analysis, comparison, and optimization of a broad range of parameter mapping/estimation methods T1, T2 1, B1+, B0, MT, diffusion coefficient 2, etc. [1] Jones et al. JMR 1996 Oct;113(1): [2] Brihuega-Moreno et al. MRM 2003 Nov;50(5):

Idea The Cramer-Rao Lower Bound (CRLB) provides a “best case” limit on the variance of an estimator of parameter – In quantitative MR, the estimator is defined by the applied pulse sequence and protocol details Typically this is used as an analysis tool, but what if we instead use it as a cost function? – We want to solve the problem: Find the protocol that gives the lowest variance estimate of θ – This is hard to solve for non-linear equations, relax it to: Find the protocol that gives the best lower bound on the variance of the estimate of θ

CRLB: Fisher Information Matrix

CRLB: How does it work?

CRLB: Computing the Jacobian Questionable accuracy Numeric differentiation Has limited the application of CRLB Difficult, tedious, and slow for multiple inputs, multiple outputs Symbolic or analytic differentiation Solves all these problems Calculation time comparable to numeric But 10 8 times more accurate Automatic differentiation

Automatic Differentiation An algorithm for generating computer programs that calculate derivatives of other functions based on their source code/graph – Uses repeated application of the chain rule – Fast and accurate to machine precision – Can differentiate extremely complicated functions – Many packages exist for Matlab, C, and Python Typically you will write the function that computes your signal equation – Then your AD package provides the derivatives essentially for free, i.e. no analysis or coding effort – YMMV depending on the language or package

AD: How does it work? Numeric: implement definition of derivative Symbolic: N-line function -> single line expression Automatic: N-line function -> M-line function – A technology to automatically augment programs with statements to compute derivatives

What does this buy us? AD gives us a way to compute CRLB or Fisher information cheaply Both in terms of coding effort and CPU time We can use this framework to find the optimal protocol design that provides the best estimate for a given θ This is usually the first question when someone invents a new mapping method Or we can explore variety of scenarios and design cases What is the optimal protocol for a range of T1s? What is the protocol that is least affected by B1+/B0 effects? What precision can be gained by using complex data instead of magnitude data? Which of these different mapping methods is the most efficient at estimating θ?

Application To vet this approach and my code, which is ultimately intended for open-source use – Went back to analyze a number of classic designs to see if I could reproduce them Diffusion coefficient, VFA/DESPOT1, DESPOT2 In all cases, the answer was yes The DESPOT2-cast was particularly interesting

Steady-State Relaxometry with bSSFP and SPGR (aka DESPOT2) DESPOT2 seeks to estimate T1 and T2 1.First do VFA/DESPOT1 with a pair of SPGR images at ~0.71 signal level to get a T1 map 2.Then acquire a pair of π-phase-cycled bSSFP images at ~0.71 signal level and back out T2, knowing T1 from step 1 – Constraints: Phase-cycling is π Acquisition time is equal between the angles in a pair, only optimize the relative time between the SPGR and bSSFP parts of the experiment

Methods Find α’s and acquisition time fraction Minimize the sum of the coefficient of variations (CoV) of T1 and T2: σ T1 /T1+σ T2 /T2 – Where σ was provided by the CRLB which uses the Jacobians of the SPGR and bSSFP signal equations – Exhaustively search for the # SPGR and bSSFP images – Fix the total acquisition time so we find the best protocol per unit time Optimize for representative T1 and T2 values of WM and GM at 3T – WM (T1=1100ms, T2=60ms) and GM (T1=1645ms, T2=85ms)

Solver Used a general solver for this problem – This is the weakest part of the framework as the solver sometimes fails to converge or encounters inversion errors – May be improved with insights from optimal design theory, e.g. using determinant of F instead of actual CRLB

Results To within 5 decimal places, framework says to acquire the pairs of flip angles at 1/√2 of the max signal for SPGR and π-phase-cycled bSSFP – Compared to the original literature estimate of 0.71 (Deoni 2003) – I later found that 1/√2 was shown to be the analytically correct solution (Schabel and Morell 2009) ~75% of time should be spent on SPGRs and T1 map Optimal DESPOT2 protocols achieve a sum of T1 and T2 CoVs of 45.7 and 53.0 for WM and GM

Unrestricted Problem Can we do better? What if we remove the constraints from the DESPOT2 acquisition? – Allow any combination of SPGRs and bSSFPs to be used, forget about the 2-stage estimate of T1 then T2 and reconstruct with a joint nonlinear fit – Allow any phase-cycle to be used – Allow the acquisition time for any image to be freely adjustable

PCVFA Solution The 0 and π phase cycles should be collected, each with a pair of flip angles that attain 1/√2 of the max signal for that phase-cycle Equal time fraction should be devoted to each of the 4 images Achieves a sum of CoVs of 20.8 and 21.6 for WM and GM – More than 2.1x improvement over DESPOT2

Performance of Optimal GM Protocol over a Range of Tissues DESPOT2PCVFA

Experimental Validation In progress with silicone oil phantom at 3T to mitigate B1 inhomogeneity – Harder to acquire DESPOT2 optimal sequence due to time difference between SPGR and bSSFP parts – Fractional angles by adjusting ia_rf1, <1 deg. Needed for PCVFA – Had difficulties doing linear and nonlinear fit for DESPOT2 and PCVFA with acquired data Maybe made an error in the acquisition, should try again

One step further What if we used complex images for the reconstruction? > – Another 2x gain?

Complex PCVFA vs. Magn. PCVFA

Summary This framework gives the ability to study complex pulse sequences and their efficacy analytically – A way to formally compare and explore potential mapping methods Gets closer to answering the question of, what is the best way of mapping X? – But may not include considerations like immunity to B1+ inhomogeneity or ease of fitting

Next Steps Robust protocol design – Have been looking at B0 robustness, where cost function evaluates CRLB at many B0 values and try to optimize the worst case estimate of θ among them – B1+ robustness Broad comparison of efficiency of mapping methods – T1 or B1+ mapping is probably a good candidate mcDESPOT and MWF MPnRAGE, what should n be? Analyzing bloch simulations, MRF?