HELSINKI UNIVERSITY OF TECHNOLOGY LABORATORY OF COMPUTER AND

HELSINKI UNIVERSITY OF TECHNOLOGY LABORATORY OF COMPUTER AND INFORMATION SCIENCE NEURAL NETWORKS RESEACH CENTRE Variability of Independent Components in functional Magnetic Resonance Imaging Jarkko Ylipaavalniemi 7. 3. 2005

Introduction Background § Functional Brain Imaging Becoming Mainstream § Standard Analysis Method Very Limited § Great Results with Independent Component Analysis § But Is ICA Reliable? § Bootstrapping, Icasso and Other Methods § Previous Work with Biomedical Data Purpose of the Work § Method for Consistent ICA of f. MRI Data § Tools for Processing and Visualizing f. MRI Data Jarkko Ylipaavalniemi 2

Brain Imaging (f. MRI) Hemodynamic Responses § Blood Oxygenation Dependent Signal § Complex and Delayed Relation to Electrical Activity Standard Analysis of f. MRI Sequences § Realign, Smooth and Normalize Volumes § Statistical Parametric Mapping (SPM) Jarkko Ylipaavalniemi 3

Independent Component Analysis (ICA) Application to f. MRI Data § Independence Seems Natural § Blind Method § Spatial Sequence § Temporal Information Important Jarkko Ylipaavalniemi 4

Independent Component Analysis (ICA) Continues Fast. ICA Algorithm § Fast and Robust § Handles the Huge Data Estimation Errors § Initial Conditions § Stochastic Iteration and Optimizations § Data Matrix Jarkko Ylipaavalniemi 5

Variability of Independent Components Sources of Variability § Estimation Errors § Assumption of Strict Statistical Independence § Weak Signal-to-Noise Ratio Exploiting with Multiple Runs § Randomizing the Initial Conditions § Resampling the Data Matrix § Clustering the Estimated Components Jarkko Ylipaavalniemi 6

Visualization and Interpretation Lots of Multi-Modal Information § Structural MRI § Spatiotemporal Activation Patterns § Clustering Statistics § Temporal Variability § Sometimes Even Spatial Variability Superimposing All Together § Interactive User Interface § Allows Easy Interpretations from Variability § Detect Consistent Components § Constant Noise or Structured Nature § Further Analyze the Structure Jarkko Ylipaavalniemi 7

Illustrative Experiments Real f. MRI Data § Reveals True Performance § Involved 14 Subjects § Auditory Stimulation Pattern § Resulting Sequences 80 x 95 x 79 x 69 Results § About 20 Clusters x 14 Subjects § All Very Interesting Jarkko Ylipaavalniemi 8

Result Highlights Components Related to Stimulus Components Revealing Artifacts Jarkko Ylipaavalniemi 9

Result Highlights Continues Components with Strong Variability Other Interesting Components Jarkko Ylipaavalniemi 10

Result Highlights Continues Spatial Variability § Demanding Computation § More Natural in Some Cases Spatial Relations Revealed § May Identify Subspaces § Related to Physical Phenomenon? Jarkko Ylipaavalniemi 11

Conclusions The Method Works § Comparable to Other Methods § Much Faster and Easier to Use Medical Relevance § Allows Reliable ICA in Real Setups § Detection of Previously Unknown Phenomena Future Research § Incorporate Knowledge into ICA Estimation § Application of Denoising Source Separation (DSS) § Analyzing More Challenging Data Jarkko Ylipaavalniemi 12

Questions Jarkko Ylipaavalniemi 13