Date of Award


Document Type

Open Access Dissertation




Experimental Psychology

First Advisor

Chris Rorden


One of the most ubiquitous steps in neuroimaging is the normalization of brain images. The process of normalization attempts to match any given brain to a standardized template image (e.g. the MNI 152 image). However, clinical images such as those from stroke participants present many challenges when we attempt to warp them to the space of template images, which are typically representative of neurologically healthy individuals. Many software packages exist to facilitate normalization of brain images, but most have limited options available to compensate for brain injury, which is often disruptive to these algorithms. Of the injury compensation methods that do exist, they are varied across software packages. The current study aimed to assess the contemporary methods available in state of the software commonly used across the field. Specifically, we assessed SPM12’s new tissue filling procedure on masked clinical images, and LINDA, a fully automated lesion segmentation algorithm combined with ANTs normalization. Across normalization methods, we compared each software package’s default injury compensation strategy to the nonstandard enantiomorphic lesion healing procedure. We created an artificial dataset of more than 10,000 images representing stroke related injury, and assessed each normalization method (SPM’s unified segmentation, DARTEL, ANTs) on multiple performance metrics. Overall, we found that the optimal injury compensation strategy for clinical images varied by the normalization method used, and the metric it was evaluated on. Finally, we present evidence of each vi normalization method and brain injury compensation technique’s effect on predicting behavior deficits from brain injury using support vector regression. Our results show that prediction accuracy (and error) can be affected by the normalization technique used.


© 2018, Taylor Hanayik

Included in

Psychology Commons