A Semy–Analytic Bayesian Approach for Multiple Static Dipoles Estimation from a Time Series of MEG Data Sara Sommariva Dipartimento di Matematica, Universita` degli Studi di Genova sommariva@dima.unige.it 1 / 16 1 MEG Inverse Problem and Source Model 2 Statistical Model 3 A Novel Approach: Semi–Analytic SMC 4 Results sommariva@dima.unige.it 2 / 16 MEG Inverse Problem and Source Model MEG Inverse Problem The aim is to reconstruct the neural currents inside the brain, from the magnetic field recorded outside the scalp. From the mathematical point of view: inversion of the Biot–Savart operator. MEG device MEG sensors helmet sommariva@dima.unige.it 3 / 16 MEG Inverse Problem and Source Model Data: the magnetic field recorded by the Ns sensors at Nt time–points y = (y1 , · · · , yNt ) ∈ RNs Nt Butterfly plot: sommariva@dima.unige.it 4 / 16 MEG Inverse Problem and Source Model Static Multi–Dipolar Model with Time–Varying Moments Single dipole: q δ (r − r (c )) where q dipole moment r (c ) dipole position. c Points of a brain–grid {r (c )}N c =1 sommariva@dima.unige.it 5 / 16 MEG Inverse Problem and Source Model Static Multi–Dipolar Model with Time–Varying Moments Single dipole: q δ (r − r (c )) where q dipole moment r (c ) dipole position. c Points of a brain–grid {r (c )}N c =1 At each time–point t the neural currents are approximated with d X (k ) qt δ r − r (c (k ) ) k =1 where d ≤ Dmax is the number of active dipoles. sommariva@dima.unige.it 5 / 16 MEG Inverse Problem and Source Model Data: the magnetic field recorded by the Ns sensors at Nt time–points y = (y1 , · · · , yNt ) ∈ RNs Nt Unknown: number of dipoles and their parameters x = (d , c, q) = d , c (1) , . . . , c (d ) , q1(1) , . . . q1(d ) , . . . , qN(1t ) , . . . , qN(dt ) From Biot-Savart equation (conditionally linear model): y = G (d , c)q + e where G (d , c) is the leadfield matrix e is the noise affecting the measurements. sommariva@dima.unige.it 6 / 16 Statistical Model Bayesian Approach to Inverse Problems Data, unknown and noise are modeled with random variables. The aim becomes to approximate the posterior pdf π(x|y). sommariva@dima.unige.it 7 / 16 Statistical Model Bayesian Approach to Inverse Problems Data, unknown and noise are modeled with random variables. The aim becomes to approximate the posterior pdf π(x|y). From Bayes theorem: π(x|y) = π(x)π(y|x) π(y) where π(x) is the prior pdf: embodies all information about the unknowns available before the data are recorded; π(y|x) is the likelihood function: embodies the forward problem and the noise model. π(y) is a normalizing constant. sommariva@dima.unige.it 7 / 16 Statistical Model We assume: Prior: π(x) = π(d )π(c|d )π(q|d , c) 1 number of dipoles: π(d ) Poisson with small, fixed parameter; 2 dipole positions: π(c|d ) = π(c (1) , . . . , c (d ) |d ) Uniform on the brain–grid; 3 dipole moments (time–varying): (1) (d ) π(q|d , c) = π(q1 , . . . , qNt |d , c) = N (q; 0, σq 2 I). Likelihood: π(y|x) = πe (y − G (d , c)q) = N (y − G (d , c)q; 0, σe 2 I) sommariva@dima.unige.it 8 / 16 A Novel Approach: Semi–Analytic SMC Novel Approach: Analytic Results Under the previous assumption we obtain an analytic expression for The marginal likelihood: π(y|d , c) = N (y; 0, Γ(d , c)) The conditional posterior over the dipole moments: π(q|y, d , c) = N (q; σ2q G (d , c)T Γ(d , c)−1 y, σ2q I − σ2q G (d , c)T Γ(d , c)−1 G (d , c) G (d , c)) where Γ(d , c) = σ2q G (d , c)G (d , c)T + σ2e I sommariva@dima.unige.it 9 / 16 A Novel Approach: Semi–Analytic SMC Novel Approach: saSMC Algorithm We propose the following two–step algorithm [Sommariva and Sorrentino, submitted]. π(x|y) = π(d , c|y)π(q|y, d , c) First step: approximate π(d , c|y) through Sequential Monte Carlo (SMC) samplers [Del Moral et al., 2006]: π(d , c|y) ' I X W i δ(d − D i )δ(c − Ci ). i =1 Second step: for all the particles {(D i , Ci )}Ii=1 obtained at the previous step we analytically compute π(q|y, D i , Ci ). sommariva@dima.unige.it 10 / 16 A Novel Approach: Semi–Analytic SMC First step: approximate π(d , c|y) through Sequential Monte Carlo (SMC) samplers [Del Moral et al., 2006]. Sample sequentially from πn (d , c|y) ∝ π(d , c)π(y|d , c)αn , 0 = α1 < · · · < αN = 1 That is Initialization: sample the first set of I particles {(D1i , Ci1 )}Ii=1 from π(d , c). Assign uniform weights W1i = 1I . Evolution: for each particle, sample (Dni +1 , Cin+1 ) from Kn+1 ((Dni , Cin ), · ), Kn+1 (·, ·) being a Markov kernel πn+1 –invariant (Reversible Jump). Assign weight Wni +1 = Wni π(y |Dni , Cin )αn+1 −αn . sommariva@dima.unige.it 11 / 16 A Novel Approach: Semi–Analytic SMC Connection with Previous Works Our work improves the one in [Sorrentino et al., 2014] 1 in which the SMC procedure is used to approximate the whole posterior π(x|y) (full SMC). Main advantages: we can give in input time–window without increasing the computational cost; the number of particles I being equal, reduction of the Monte Carlo variance. 1 Talk by Alberto Sorrentino on Wednesday 9th July at h. 11.00 MS 8. sommariva@dima.unige.it 12 / 16 Results Simulated Experiment Forward Problem: sommariva@dima.unige.it 13 / 16 Results Simulated Experiment Forward Problem: sommariva@dima.unige.it 13 / 16 Results Simulated Experiment Forward Problem: sommariva@dima.unige.it 13 / 16 Results Simulated Experiment Forward Problem: sommariva@dima.unige.it 13 / 16 Results Simulated Experiment Forward Problem: sommariva@dima.unige.it 13 / 16 Results Simulated Experiment Forward Problem: sommariva@dima.unige.it Source Reconstruction: 13 / 16 Results Simulated Experiment Forward Problem: sommariva@dima.unige.it Source Reconstruction: 13 / 16 Results Real Data: SEF Data sommariva@dima.unige.it 14 / 16 Results Real Data: SEF Data SA SMC: full SMC: dSPM: sommariva@dima.unige.it 14 / 16 Results Real Data: SEF Data SA SMC: full SMC: dSPM: sommariva@dima.unige.it 14 / 16 Results Future Work and Acknowledgements More accurate/automatic choice of the parameters, especially σq . Application to other real data (good results with EEG auditory data). Application to other inverse problems with conditionally linear models. Joint work with Alberto Sorrentino, Dipartimento di Matematica, Universita` di Genova. sommariva@dima.unige.it 15 / 16 Results Reference Del Moral, P., Doucet, A., and Jasra, A. (2006). Sequential Monte Carlo samplers. Journal of the Royal Statistical Society B, 68:411–436. Sommariva, S. and Sorrentino, A. (submitted). Sequential monte carlo samplers for semi–linear inverse problems and application to magnetoencephalography. Inverse Problems. Sorrentino, A., Luria, G., and Aramini, R. (2014). Bayesian multi-dipole modeling of a single topography in meg by adaptive sequential monte-carlo samplers. Inverse Problems, 30:045010. sommariva@dima.unige.it 16 / 16