From Pixels to Brain Networks: Modeling Brain Connectivity and Its Changes in Disease Polina Golland MIT Computer Science and Artificial Intelligence Laboratory Joint work with Archana Venkataraman C.-F. Westin Marek Kubicki
Observing Brain Networks DWI scans: diffusion of water along fibers Tractography estimates white matter bundles Rest state fmri: activation measures Correlation as a measure of co-activation Today: region-based connectivity Question: what can we say about brain networks and how the disease affects them?
Connectivity Studies Univariate tests on each connection separately Result: a list of connections Mulitmodal studies: Fusion as post-processing Network analysis: Compute statistics and identify the type of network» E.g., small world How do we go from these answers to meaningful clinical statements?
Our Solution Latent network model Group-wise latent connectivity template Image-based measures are noisy observations Joint inference from anatomical and functional connectivity data Region-centric inference Foci of disease
Latent Model of Connectivity Template D 1 ij B 1 ij 1 A ij F ij D L ij A - anatomical connectivity (binary); D - DWI observation F - functional connectivity (trinary); B - fmri observation B L ij L
Data Likelihood (DWI) Template ρ 0 P (D l ij A ij = 0; ) D 1 ij B 1 ij 1 A ij F ij D L ij B L ij L
Data Likelihood (DWI) P (D l ij A ij = 1; ) D 1 ij Template ρ 1 B 1 ij 1 A ij F ij D L ij B L ij L
Data Likelihood (fmri) P (B l ij A ij =0,F ij = 1; ) Template D 1 ij B 1 ij 1 A ij µ 0, 1 F ij D L ij B L ij L
Data Likelihood (fmri) P (B l ij A ij =0,F ij = 0; ) Template D 1 ij B 1 ij 1 A ij µ 00 F ij D L ij B L ij L
Data Likelihood (fmri) P (B l ij A ij =0,F ij = 1; ) Template D 1 ij B 1 ij 1 A ij µ 01 F ij D L ij B L ij L
Population Differences Control Template Clinical Template A ij Ā ij F ij F ij
Complete Generative Model Control population Shared Data Likelihood Clinical population This is clustering!
Application to Schizophrenia 19 Chronic Schizophrenia Patients, 19 Controls T1 scans segmented into 77 Regions (Freesurfer) DWI, fmri standard preprocessing Region-Wise DWI Measures Region-Wise fmri Correlations
Functional Differences P L Functional connections for which ˆ f ij > 0.75 Reduced connectivity between parietal/posterior cingulate and temporal lobe R Reduced connectivity in schizophrenia Increased connectivity in schizophrenia A Increased connectivity between parietal/posterior cingulate and frontal lobe Affects perception of reality
Anatomical Differences R P Anatomical connections for which ˆ a ij > 0.75 Implicate the cuneus, superior temporal gyrus and hippocampus A L Inter-hemispheric connection is an artifact of tractography Reduced connectivity in schizophrenia Increased connectivity in schizophrenia
From Connectivity Measures to Region-Centric Conclusions The answer comes as a list of connections P L All other knowledge and discussion is about regions Hard to integrate connectivity findings No good way to design experiments for connections R A Our solution: model region-based connectivity changes Sub-network of abnormal connections
Region-Based Model R 9 R 8 R 2 R 10 R 7 R 3 R 1 R 11 R 6 R 4 R 5 R 12
Region-Based Model (cont d) R 9 R 8 R 2 R 10 R 7 R 3 R 1 R 11 R 6 R 4 R 5 R 12
Complete Generative Model Latent Structure Shared Data Likelihood
Detected Regions L STG R STG R PCC
Abnormal Sub-Network P L R A Reduced connectivity in schizophrenia Increased connectivity in schizophrenia
Prior Parameter Sweep
Conclusions Latent network model of connectivity Joint inference from all available data and over all connections From a list of connections to a set of regions Disease foci Affected sub-networks Clinical applications Today: schizophrenia Variant of this model: epilepsy Support: NAMIC: National Alliance for Medical Image Analysis NAC: Neuroimaging Analysis Center