Blind Separation of sources in function MRI Sequences Presented By: Eldad Klaiman Limor Goldenberg Supervised By: Michael & Alex Bronstein Dr. Michael Zibulevsky The Kasher Contest - In memory of Yehoraz Kasher The Problem functional MRI: Blind Source Separation • • • • • Important tool for studying the human brain activity. High spatial resolution, flexibility, harmlessness – made it popular. The BOLD technique: produce an image of the blood oxygenation level throughout the brain. A sequence of scans is in a short period of time, when the subject is asked to perform some task. High oxygenation levels represent high activity of the brain regions responsible for the task. • • • • Linear mixture of independent sources No a priori information is known about their properties. “Blind Source Separation" = the problem of separating such sources. There exist powerful tools to solve it. Focus on the approach of sparse representations, which has proved its advantages in different works in the field. fMRI-BSS Model Sources Separated Sources Mixtures S1 M1’ M1 S2 Pre Processing fMRI Mn Sn m As S1’ Separation Mn’ -Noise removal -Identify Background -Sparse Representation Sn’ ŝ A 1 m fMRI Simulation • Background – Brain Image. (x x ) ( y y ) S ( x , y ) exp{ } • Spatial Function: • Hemodynamics: H (t ) t exp{ (t t ) } 2 2 0 0 i s 2 0.3 i t 0 • Gaussian Noise M (k ) Si ( x, y ) H i (t ) Background Noise i fMRI simulator GUI fMRI Simulator - Results “fMRI” frames Hemodynamics Preprocessing – Sparse Representations • Wavelet Packets is used to create sparse images. • “Best” Node is selected by sparseness Criteria • Scatter plot of resulting images: chasing the illusive “X” Geometric Separation • Clustering - FCM. • Angle Histogram. Separation Example Source #1: 3 spatial components Source #2: 2 spatial components Issues Encountered • Preprocessing : Zero-mean, LPF, etc. • Sparseness Criteria : Shannon entropy selected. • Stability / Parametric Sensitivity : thresholds. Principal Component Analysis • Problems of high order: more mixtures than sources • Problem dimension reduced using PCA PRINCOMP( ) PCA Revelations (1) Background Separated from activity sources (2) No need to know the exact number of sources. ICA - Infomax • Artificial Neural Network Viewpoint, maximize output Entropy. • InfoMax ICA Matlab Toolbox: (courtesy of Scott Makeig & Co.) • Preliminary Results can be obtained without mixture preprocessing. ICA Separation Example ICA Notes • Sign and Order limitations. • Improved robustness and quality, compared to geometric separation. • In most cases, the sparse representation improved the quality of separation. Application on the Real Thing Real fMRI Issues False Artifact Sources – created due to head movement, Noise. Background separated from activity sources. Conclusions • Achieved good results by geometric and ICA separation. • ICA – robustness, quality. • PCA – model selection, added values. • Potential as fMRI analysis tool. – Quick, low cost. – Exact knowledge of simulation flow - not needed. – Not relying on high time resolution. Further Progress • A new horizon for fMRI-ICA academic research and projects. • A “friendly” and enhanced fMRI-ICA application was developed for simple, useroriented application of algorithm. • Experimental application of the separation algorithm on LORETA (EEG-CAT). Thanks to… • • • • • Nethaniel’s Brain Dr. Michael Zibulevsky Johanan Erez and the Lab team Michael & Alex Bronstein Anat Grinfeld Sparseness Criteria • • • • x 1 xi L1: % x : xi 0 L0: Shannon Entropy: H ( x) xi log xi Clusters: q max( d ) min( ) Back