Machine Learning Classification of Traumatic Brain Injury Patients Versus Healthy Controls Using Arterial Spin Labeled Perfusion MRI

Traumatic brain injury (TBI) is one of the most common causes of death and disability worldwide, yet accurate in vivo detection of TBI neuropathology remains challenging due to complexities in the structural and functional changes observed post-injury as well as limitations in conventional neuroimaging modalities. Although advanced neuroimaging techniques such as arterial spin labeling (ASL) can noninvasively assess cerebral blood flow (CBF) changes observed post-injury, this technique is underutilized in TBI research partly due to the low signal-to-noise-ratio (SNR) inherent in ASL imaging. The aim of the current study is to examine the use of machine learning, specifically a Support Vector Machine (SVM) classifier, in discriminating between healthy controls (n=35) and TBI patients (n=42) using ASL-generated CBF data 3 months post-injury. Identification of the regions of interest (ROIs) most predictive of TBI is also explored as part of this aim. Furthermore, several ASL outlier cleaning methods, such as the Structural Correlation- Based Outlier REjection (SCORE) and prior-guided, slice-wise adaptive outlier cleaning (PAOCSL) algorithms, are examined in relation to improving the SNR and SVM performance. While the classification models tested did not reach statistically significant performance levels, the results were in the direction suggesting that more sophisticated outlier cleaning methods can improve classification accuracy. Potential explanations of the observed low classification accuracy and the implications of our findings on future research are discussed