The previously reported diagnostic performance of diffusion-tensor imaging (DTI) for mild traumatic brain injuries (TBIs) could be called into question due to bias introduced by the pothole approach to neuroimaging data, according to a study published online March 14 by Radiology.

DTI has recently shown promise for detecting the subtle effects of mild TBIs. However, diagnostic performance characteristics of DTI for mild TBI are still unknown. “Pothole” analysis allows for examination of the white matter in mild TBI and offers an alternative approach to region of interest analysis.

“In this technique, the white matter fractional anisotropy (FA) value at each voxel is transformed into a z statistic based on the mean and the standard deviation of FA in a reference population,” wrote lead author Richard Watts, PhD, of the University of Vermont in Burlington, and colleagues. “White matter potholes are defined as clusters of voxels in which the FA z statistic is below some threshold. We similarly define molehills as clusters of voxels with increased FA. Additional constraints, such as defining a minimum cluster size, may be applied. The summary statistic may be the total number of such clusters or the total volume of clusters.”

Watts and colleagues aimed to identify the extent of bias in a clinical study that included “pothole” analysis of DTI data used to quantify white matter lesion load in mild TBIs.

DTI data were ascertained from 20 patients who were admitted to the emergency department with mild TBIs and 16 control subjects. Standard analysis revealed 102.5 potholes in mild TBI and 50.6 in the control group, which was a significant difference. Using leave-one-out cross-validation, repeat analysis showed a decrease in the apparent difference in potholes between the two groups. The bias demonstrated its ability to decrease the voxelwise false-positive rate by at least 30 percent in the control group.

“Our findings are important because they suggest that the diagnostic utility of DTI ‘pothole’ analysis as reported in the literature may be overly optimistic, owing to bias in the analysis that effectively decreases the number of both molehills and potholes in the control group; this bias may be minimized in future studies by using the corrected z statistic threshold or the leave-one-out method,” wrote Watts and colleagues.

The authors advise that caution be taken when implementing these techniques into clinical practice.