Artifact Classification of fMRI Networks
1 like160 views
This document outlines a project to develop standards for defining functional brain networks using spatial and temporal features extracted from fMRI data. The goal is to distinguish true neural networks from noise and to classify neuropsychiatric disorders based on patterns of functional networks. Over 1,500 independent components were analyzed to identify 124 features that can predict whether a component represents a true network or noise, with 86.75% accuracy. Further work will focus on defining subnetworks within main networks and investigating patterns of subnetwork activity to distinguish disorders. Developing standards for networks will allow automated network labeling and filtering of fMRI data to advance diagnosis and subtyping of neurological disorders.
Artifact Classification of fMRI Networks
- 1. 105 Temporal Features signal metrics, peaks, kurtosis, skewedness, entropy, amplitudes, power bands, HPSD, auto correlation etc. PreprocessingFunctional MRI Spatial and Temporal Features of fMRI Networks to Distinguish Real Networks from Noise Preliminary work to use patterns of functional networks to classify neuropsychiatric disorder V. Sochat, Rubin Lab, Stanford University School of Medicine, Stanford CA Introduction • Independent Component Analysis ICA is a data-driven method to decompose functional neuroimaging data into brain networks. • The decomposed independent components encompass a mix of true neural signal, machine artifact, motion, and physiological noise that are typically visually distinguished. • Neurological disorders are beginning to be understood based on aberrant brain structure and function on the single network level. • No methods exist for identifying patterns across all networks to distinguish disorder. What Does the Data Look Like? How do we Define a Standard? Lasso L1 Constrained linear regression selects 124 features to distinguish real from noisy components (N=1518) with a cross validation accuracy of .8675. Supported by Microsoft Research, NSF Stanford Graduate Fellowship Can We Predict Component Type from Features? Meeting with MIND Institute 9/24/2012 to finalize standard development to allow for robust computation of functional network fingerprints. Why Should I Care? What is a network? There are no standards or definitions of functional brain networks beyond expert opinion. What is a subnetwork? Overdetermined ICA produces “subnetworks,” but we do not completely understand how they match to main networks. How to Rx Disorder? Diagnosis of neuro- psychiatric disorder with the DSM is categorical, checklist based, and terrible. We need methods to establish standards for networks and noise. Computational definition of subnetworks matching to main networks will allow for investigation of patterns of neural activity within main networks. Biomarkers from functional imaging can drive diagnosis of disorder and subtyping. 1. Standard and Features 2. Subnetwork Definition 3. Classify Disorder This Project: Spatial and temporal features to distinguish real from noisy components Next Stage of Work: Features of subnetworks and matching to main networks Big Picture: Patterns of functional brain networks for classification of neuropsychiatric disorder unsolvedproblemsobjective Realign / Reslice Motion Correction Segmentation Smoothing Filtering Normalization ICA n x m n x n n x m Contact vsochat@stanford.edu What Features Define the Networks? 1. Outline: database of “known” networks scattered in literature. Currently identification is done manually. 1. component type (real, noise, etc.) 2. network type (motor, visual, default mode network) 3. network name (“precuneus posterior cingulate”) 4. intuitive name (“the tie fighter”) 135 Spatial Features Regional activation, matter types, kurtosis, entropy, skewedness, degree of clustering Selected Features “Eyeballs” Component Perfect_total_activation_in_GM Percent_total_activation_in_WM Olfactory_R Skewedness of IC distribution Avg_distance_btw_10_local_max Spatial Entropy of IC distribution Percent_total_activation_in_eyeballs LASSO cva: 0.9841 2. Teach send database and networks to experts 3. Label: experts use annotation tool to label networks 4. Fingerprint: define networks based on pattern of features Good / Bad Network X / Not Network X 42 Networks 11 2 3 4 34 Thank You Rebecca Sawyer and Kaustubph Supekar, Signal Processing Daniel Rubin: Best Advisor Ever! 24 healthy control 29 schizophrenia Network fingerprints will allow for automatic labeling and filtering of single subject ICA. For the first time, subnetworks can be assigned meaningful labels, the main networks they match to. Different patterns of these subnetworks within the space of a main network will distinguish disorder. These methods will be extended to other imaging data Bad Good Bad 740 121 Good 78 579 N = 1518 Networks spatial maps and timecourses Please see research journal for preliminary match work