Jeremiah Flannery: Feature Extraction and Scaling Analysis of rs-fMRI Data: Examining Preferred Methods and Parcellations

BCCN Berlin / Technische Universität Berlin

Abstract

 

 

Using over 20k resting state f-MRI paired with patient data, I extracted features related to functional connectivity, graph analysis, brain anatomy, and regional signal strength with different granularities of three cortical atlases. For seven qualitatively similar groups of variables, I performed analysis with a linear and nonlinear Lasso Regression approach, comparing this result with those from a lightweight CNN approach. Results showed that for several variable sets, the extracted signals do not provide significant predictive gains. For age and sex related variables, the Brainnetome atlas gives the best results, but the ICA atlas gives nearly as accurate results while having less granularity and using less computational resources to extract features and perform analysis.

 

Additional information:

Master thesis defense

 

Organized by:

Prof. Dr. Kerstin Ritter & Prof. Dr. Med. Dr. Henrik Walter