Multi-site FMRI data normalization to reduce inter-scanner variability of activation

Abstract

Since the introduction of functional Magnetic Resonance Imaging (fMRI), a non invasive technique for mapping the brain activity, there has been an exponential growth in the brain mapping studies. Recently, multi-center fMRI studies have become more prevalent as they help accumulate significant number of subjects to increase the statis tical power of data analyses. However, the seemingly ambitious gain is hindered by the fact that diversity in acquisition and analysis methods of different centers has significant effect on the imaging results. Inspired by the use of correction functions in the domain of structural MRI, current work proposes the use of machine learning techniques to esti mate mapping functions to normalize the activation parameters observed at a particular site to a chosen reference site. Unlike previous studies that focused on normalizing acti vation related parameters such as signal-to-fluctuation noise ratio (SFNR), smoothness, etc., the main contribution of this work is to enhance the comparability of activation patterns across sites. This research work explores the use of (a) Linear regression with robust statistics, (b) Machine learning based non-linear regression and (c) Transfer learning algorithms to reduce the inter-scanner variability of activation. Additionally, a threshold adjust ment approach based on theoretical relationship between acquisition and image quality parameters is also investigated. Extensive experiments conducted using the FBIRN Phase 1 Traveling Subjects dataset (specifically designed to help assess test-retest and between-site reliability) demonstrate the potential of machine learning based algorithms to generate correction functions. Both quantitative and qualitative results indicate sig nificant reduction in spurious activations and more importantly, enhancement of the genuine activation clusters. Group level ROI based analysis reveals changes in activa tion pattern of clusters that are consistent with their role in cognitive function. Further more, as the mapping functions exhibit the tendency to induce sensitivity to the regions associated with the task they can help identify small but significant activations which could otherwise be lost due to population based inferences across centers.