Bayesian model selection under spatial uncertainty for functional imaging studies Alexis Roche CIBM-Siemens, Lausanne, Switzerland Neurospin, CEA, Paris, France alexis.roche@gmail.com 6/14/2012 Where's your signal? OHBM'12 workshop 1 Mass univariate model limitations 1. No control over sensitivity N/S 6/14/2012 Where's your signal? OHBM'12 workshop 2 Mass univariate model limitations 2. Limited localization power Significant clusters at FWER=5% depending on cluster forming threshold 6/14/2012 Where's your signal? OHBM'12 workshop 3 Mass univariate model limitations 3. Specificity reduced by registration errors Group t-map without smoothing 6/14/2012 Group t-map with 6mm smoothing Where's your signal? OHBM'12 workshop 4 Proposed approach • Bayesian formulation (addresses 1) • Regional activity is modeled using a predefined parcellation (addresses 2) • Explicit registration error model (addresses 3) 6/14/2012 Where's your signal? OHBM'12 workshop 5 Statistical model Registration error fields Individual effect maps 𝑢1 𝛽1 fMRI data 𝑌1 𝑢2 Parcel level group activity Voxel level group activity 𝜔 𝜇 𝜇𝑣 = 𝜔 + 𝜉𝑣 𝛽2 𝑌2 𝑢3 𝛽3 𝑌3 𝑢𝑛 𝛽𝑖 = 𝑋𝐺 𝜇 + 𝜁𝑖 𝛽𝑛 𝑌𝑛 𝑌𝑖 = 𝑋𝑖 𝛽𝑖 + 𝜀𝑖 ∘ 𝑢𝑖 −1 6/14/2012 Where's your signal? OHBM'12 workshop 6 Inference via belief propagation (BP) 𝑢1 𝑌1 𝛽1 𝑢𝑖 𝜔 6/14/2012 𝜇 𝛽𝑖 Where's your signal? OHBM'12 workshop 𝑌𝑖 7 Approximate BP 𝑢1 𝑌1 𝛽1 𝑢𝑖 𝜔 6/14/2012 𝜇 𝛽𝑖 Where's your signal? OHBM'12 workshop 𝑌𝑖 8 Approximation schemes • Monte Carlo – See (Keller et al, Statistica Sinica 08, MICCAI 09) – Either slow or inaccurate • Variational – Approximate messages by factorized Gaussian distributions – Amounts to sequential I-projections – Fast and numerically stable 6/14/2012 Where's your signal? OHBM'12 workshop 9 Variational BP step by step 1. Run usual first-level GLM without spatial smoothing 6/14/2012 Where's your signal? OHBM'12 workshop 10 Variational BP step by step 2. Smooth contrast and variance images via Raw 6/14/2012 2 𝐸(𝛽𝑖 ) = 𝐺 ∗ 𝛽𝑖 𝑉 𝛽𝑖 = 𝐺 ∗ 𝜎𝑖 2 + 𝐺 ∗ 𝛽𝑖 − (𝐺 ∗ 𝛽𝑖 )2 Contrast image Contrast variance image Smoothed Raw Where's your signal? OHBM'12 workshop Smoothed 11 Variational BP step by step 2. Smooth contrast and variance images via 2 𝑉 𝛽𝑖 = 𝐺 ∗ 𝜎𝑖 2 + 𝐺 ∗ 𝛽𝑖 − (𝐺 ∗ 𝛽𝑖 )2 𝐸(𝛽𝑖 ) = 𝐺 ∗ 𝛽𝑖 noise Contrast image Raw 6/14/2012 spatial uncertainty Contrast variance image Smoothed Raw Where's your signal? OHBM'12 workshop Smoothed 12 Variance estimate BP 6/14/2012 SPM Where's your signal? OHBM'12 workshop 13 Variational BP step by step 3. Estimate voxel level group parameters using mixed effects variational Bayes algorithm Group mean 6/14/2012 Group variance Where's your signal? OHBM'12 workshop 14 Variational BP step by step 4. Estimate parcel level group parameters using mixed effects variational Bayes algorithm Effect size 6/14/2012 Probability of positive effect Where's your signal? OHBM'12 workshop 15 Method comparison “Speaker” effect BP 6/14/2012 Using SPM effect maps, assuming zero variance Where's your signal? OHBM'12 workshop Using SPM effect and variance maps 16 Method comparison “Sentence” effect BP 6/14/2012 Using SPM effect maps, assuming zero variance Where's your signal? OHBM'12 workshop Using SPM effect and variance maps 17 Conclusion • Bayesian model selection may mitigate sensitivity/specificity issues of mass-univariate inference • Fast and numerically stable implementation • Available in NiPy (www.nipy.org), development version 6/14/2012 Where's your signal? OHBM'12 workshop 18