How Does the Accuracy of Intracranial Volume Measurements Affect Normalized Brain Volumes? Sample Size Estimates Based on 966 Subjects from the HUNT MRI Cohort

Accuracy of the Automated Methods versus Manual Segmentation

The automatic methods for estimating the ICV, which we evaluated, produced ICV estimates that closely matched those of manual segmentation. The ICV FreeSurfer estimates had a bias that was larger than the other methods, and FreeSurfer consistently underestimated the ICV. A possible cause is that the default scaling factor in FreeSurfer is not optimal for the present study, or that differences between the segmentation protocol for the images on which the scaling factor was optimized and that of the present study could account for the bias.

Optimizing the FreeSurfer scaling factor improved the ICV estimates. A drawback is that one must have a sufficiently large set of images with manually derived ICV measurements to compute an optimized scale factor. Future studies could determine whether the variation in the optimal scaling parameter is primarily determined by the scanner parameters or by the study population.

The RBM method was the most accurate for estimating the ICV. We also found that in comparison with the original implementation of the RBM method, the “new segment” algorithm in SPM improved the accuracy of the RBM method. (See On-line Table 3 for a summary of the accuracy of the original RBM method.) A disadvantage of the RBM method, however, is that one must set an empiric threshold for calculating the ICV. Therefore, the accuracy of the RBM method is dependent on the threshold. This dependency is illustrated in On-line Table 3 showing the accuracy of the RBM method with the optimized threshold and with the threshold recommended by the authors of the RBM method.⁴ Using the nonoptimized threshold renders the RBM method less accurate than the ARBM method. We also found that a visual determination of the threshold was difficult because it varied among different raters. Optimizing the threshold against the manual segmentation result avoided this problem but is impractical in many instances because it necessitates manual measurements.

The ARBM method attempts to alleviate the drawback of using a threshold. An ICV mask must still be drawn in template space, but it needs to be done only once. The ARBM approach was, however, less accurate than the RBM method but more accurate than the FreeSurfer methods. The ICV estimates with the ARBM method may be robust over different field strengths, similar to those with the RBM method,⁴ because the 2 methods only differ in how the brain masks are thresholded.

Multispectral input clearly improved the accuracy of the RBM method, suggesting that the transformation from template space to native space is more accurate when using T1 and T2 images as input compared with using only T1 images. For the ARBM method, however, multispectral input resulted in a slight underestimation of the ICV. This discrepancy in accuracy between these methods can appear puzzling because they rely on the same transformation from template space to native space. The underlying cause is that the multispectral segmentation, on average, generates a slightly smaller volume in native space than the segmentation based on T1 images only. The bias is adjusted during the optimization of the threshold in the RBM method because the optimized threshold for RBM multi is 0.29 compared with 0.34 for the RBM method (a lower threshold results in a larger ICV mask). For the ARBM method, the ICV is fully determined by the transformation to native space; therefore, there is an increase in the mean volume difference for the ARBM multi method. The bias in the ARBM multi estimates is mainly a concern when using ratio correction. For residual correction, the ARBM multi method would still be preferable over the ARBM method because the multispectral segmentation reduces the variance in the volume estimates compared with the T1-only ARBM. This outcome is reflected in the slight decrease in sample size estimates for the ARBM multi method over the ARBM method (Table 2).

Sample Size Estimates

The ICV is often used for correcting variations in regional brain volume measurements due to differences in head size. Several studies have compared the accuracy of various ICV-estimation methods,^3,4,21,22 but surprisingly few have examined the practical benefits of an accurate ICV measurement. Naively, one would expect that accurate ICV estimates would increase the statistical power of ICV-corrected regional volume measurements. Our results demonstrate that not only the choice of ICV estimate, but also the method of ICV correction can affect the statistical power. We found that residual correction resulted in only minor differences between the FreeSurfer and ARBM methods (Table 2). In fact, the FreeSurfer ICV correction generally required a smaller sample size than the ARBM-corrected volumes. This is surprising considering that the ARBM method had higher accuracy than FreeSurfer compared with manual segmentation (Table 1). However, the differences in the required sample sizes for the volume estimates can largely be explained by the strength of the correlation between the ICV and the volume measurements (Fig 3).

We found that accurate ICV estimates were more crucial for the ratio-corrected volumes, a finding that is in agreement with previous studies,^9,19 and that the ratio-correction method, unlike the residual-correction method, requires absolute agreement in the ICV estimates. When using ratio correction, we found that the more accurate ARBM ICV estimates can provide increased power compared with the FreeSurfer ICV estimates. For example, to detect a 2% difference in the hippocampus volumes requires 44 fewer subjects per group when using ARBM ICV estimates compared with FreeSurfer ICV estimates. However, the difference in the required sample size becomes smaller for larger effects (Fig 2); for medium-sized or larger effects, there are only minor differences among the methods we evaluated.