To investigate inconsistencies between common performance measures for bias field correction reported in several recent studies and propose a solution.Materials and Methods:
A set of synthetic images of a normal brain from the Montréal Simulated Brain Database (SBD) was processed using two bias field correction algorithms. The parameters of these algorithms were varied and the resulting outputs were assessed using several performance measures. Validity was estimated using Spearman rank correlation coefficient between “indirect” performance measures and the L2 norm of the difference between true and estimated bias fields. The “indirect” performance measures tested were: coefficients of variation of white matter (WM) and gray matter (GM), coefficient of joint variation. These measures were tested on bias field-corrected images that were permuted in terms of quality of WM/GM segmentation as well as the presence or absence of light smoothing.Results:
Existing indirect performance measures yielded poor validity scores, explaining the inconsistencies reported in the literature. Image noise and inappropriate inclusion of partial volume voxels and neighboring tissues were found to be contributory. Combining conservative segmentation and smoothing significantly improved validity.Conclusion:
The use of indirect performance measures in the conventional manner to guide bias field correction is unreliable. Using these metrics on lightly smoothed images with conservatively segmented tissues proved more reliable for guiding the selecting of parameters for nonuniformity correction ultimately contributing to more accurate brain segmentation.