Summary of findings
Choice of ROI method was found to significantly affect the FA values when the optic tracts were analyzed. The manual b0 and the FA-skeleton methods resulted in the highest FA values, indicating that these two methods best identified the middle of the optic tracts, where voxels are less influenced by partial volume effects. Results from the FA-skeleton method had lower variability compared to the manual b0 method for all comparisons (inter-individual, inter-scan and intra- and inter-rater analysis). This suggests that the FA-skeleton method, compared to the manual b0 method, leads to more reliable results.
Interestingly, the manual T1W and the tractography methods resulted in similar FA values, which were significantly lower than for the other two methods, indicating an inability to accurately define the middle of the small structures that are the optic tracts.
A suitable ROI method, for both clinical and research purposes, should have an ability to accurately define the structure of interest and a high reliability. Based on the results of the present study the FA-skeleton ROI method performed best according to these criteria and may be suggested for analysis of the optic tracts, and possibly of similar structures, with regard to size, shape and expected image artifacts.
Comparison of ROI methods
ROIs for DTI may be defined in images with more anatomical information, such as T1-weighted images, which are subsequently registered to diffusion space; such registration inevitably leads to image distortion to some degree [
18,
19]. Grech-Sollars et al. coregistered b0 images to high-resolution T1-weighted images, preceding registration to standard space (MNI) [
19]. They found acceptable overall FA values, however, they also found very low values for specific white matter structures, such as the optic chiasm (FA 0.18). These spuriously low values were probably due to the choice of linear registration. In the present study, whole brain coregistration between T1-weighted images and diffusion space, through different attempts of both global and local affine transformations, did not produce a sufficiently good match for the OTs, although other white matter structures were adequately matched. A likely reason for this registration failure is the size of the OTs in combination with susceptibility-artifact effects. The most successful match was achieved when a spline-based transformation was applied for registration of the T1-weighted images to the FA maps. Such a transform is able to mimic geometric effects of susceptibility artifacts. However, the FA values extracted by this method were lower than those of the manual b0 and FA-skeleton methods, suggesting a remaining slight off set after the coregistration.
In previous literature, manual ROIs in diffusion image space, such as the b = 0 map or the FA map, have been preferred for analysis of the anterior visual pathways [
5,
20,
21]. Hakulinen et al. compared two different manual ROI methods – circular and free hand – on several white matter structures [
8]. They found that FA differed depending on ROI method and, in accordance with other studies, that the variability was larger the smaller the structure [
5‐
7,
9,
10]. This effect for small structures is most likely due to the measurement procedure: the difficulty of redefining similar ROIs, leading to user-errors, and the increased risk of including boarder-zone voxels affected by partial volume effects.
More objective data extraction methods could resolve the issue of user-errors. One such proposed method is data extraction by tractography, where part of the resulting tract can be selected for data extraction. Paul et al. used probabilistic tractography to visualize and extract data from the OTs [
2]. They reported differences in DTI parameters between patients with pituitary tumors, affecting the anterior visual pathways and vision to varying degrees. However, tractographies have been shown to differ to a significant degree depending on tractography algorithm as well as on the selected parameters within the algorithm [
22]. For example, Wang et al. assessed reliability of DTI parameters by tractography-based ROIs, comparing 15 and 30 gradient sampling directions, and found significantly affected values as well as reduced variability using 30 directions [
23].
In the current study, the results by tractography-based ROIs showed a strong intra-rater agreement but a poor inter-rater reproducibility (Fig.
5). The only factors that differed between different raters were the manually chosen seed, termination and exclusion masks; although carefully anatomically defined, the differences that resulted between raters proved to have an important effect on the tractographies. Reproducibility could thus be a problem for tractography-based ROIs in small structures, which should be considered in comparison of results from different studies.
In this study, we propose a version of ROI method that aspires to bypass issues of subjectivity/user-errors, registration and tractography. The skeleton algorithm of TBSS was applied on each subject’s individual FA map, keeping the original image space. This creates an FA skeleton of the entire brain for each individual, defining the central voxels of each white matter tract [
13]. The user defined start and end of the OT, but the voxels were otherwise primarily selected by the algorithm. Compared to the other ROI methods in this study, the FA-skeleton method performed well: the mean FA was high, suggesting the middle of the tracts was successfully identified, and the variabilities across subjects and between measurements were low. Since the OTs are thin structures it could be assumed that the voxels with the highest FA include the center of the structures; the FA-skeleton normally defines two voxels per cross section and it is likely that voxels lateral of these most central voxels are affected by partial volume effect, resulting in lower FA. When performing the full TBSS procedure on all subjects, including the registration to a common space, the OTs were not identifiable. The individual FA-skeleton version of ROI method could be an alternative when comparing and extracting data from relatively small structures, and could be used for comparisons where significant anatomical differences are expected.
Reliability
Rater performance is an important factor of variability. Previous DTI reliability studies have reported significantly higher inter-rater than intra-rater variability [
24], similar to results by the tractography method in this study. However, the other three ROI methods herein show little difference between the intra- and inter-rater variability, indicating high reproducibility.
The variability due to repeat-scan was higher than the intra- and inter-rater variabilities, which is to be expected as repeated scans include the intra-rater error as well as scan-specific factors. There are several scan-specific factors that may affect measurements. The most important ones include differing slice and head positioning, where changes between scans will affect the partial volume effects and possibly also the effects of susceptibility. Varying head motion is another possible source of inter-scan variability, especially if the averaging is carried out on scans where intermediate motion may occur, which was the case in the present study [
25]. Although the intra-rater variability was lower for the tractography method compared to the other three ROI methods, the inter-scan variability for the tractography method was higher. Possibly, the tractography algorithm is more sensitive to the scan-specific changes.
When adjusting for starting slice, the variability decreased considerably (Fig.
6). The greatest decreases for the combination of intra- and inter-rater variability were seen for the manual b0 and the FA-skeleton ROI methods, which imply that the subjectivity in the identification of the start slice was an important factor of variability, whilst the ROI was otherwise well reproducible. In conclusion, small rater-dependent factors have a large impact on ROI analysis of small structures.
The Jaccard indices were in general low between ROI methods (~0.3), including between the ROI methods that showed high similarity in FA values. At first glance this may seem contradictory, however, Jaccard index only compares chosen voxels, and does not take into account the values within the voxels. The explanation for these findings could be that the ROIs in this study were chosen to be “thin” – only two voxels per coronal OT section. More than two voxels could be representative of the middle of the OT, and thus give similarly high FA values. The choice of two voxels was a compromise between reducing the risk of partial volume effect and increasing the amount of included data. In conclusion, the similarities between ROI-method results in this study could be said to be due to their ability, and similarities in ability, to identify the center of the optic tract, but not due to their ability to choose the same voxels.
Limitations
The anterior limit of the ROIs was defined as the most posterior part of the optic chiasm. Thus, a certain number of chiasm voxels will be included in
Background section (i.e., the most anterior section). In the optic chiasm there are several different fiber directions, as fibers corresponding to different parts of the visual fields are reorganized. As a consequence, the FA in such voxels will be an average of several fiber orientations. The diffusion tensor model used herein, which assumes one principal diffusion direction per voxel, may thus be poorly suited for the chiasm, including
Background section. However, the tensor model is well suited for the parallel organization of the OTs, and the focus of this study should thus be on results from
Methods and
Results sections.
Because of magnetic susceptibility effects in the region of interest in this study we used a 1.5 T scanner, instead of a 3 T, with a relatively high SENSE factor (=3.2). Both of these choices contribute to mitigate image distortion caused by magnetic susceptibility, and result also in lower SNR. Therefore increased signal averaging was used to enhance the SNR. More recent techniques for reduction of image distortions, such as zoomed acquisition, were not available at the site by the time of the study.
A single scan protocol and resolution was applied in this study. The relationships between ROI-methods reported herein may be different for other protocols, for example protocols with smaller voxels or a different number of diffusion encoding directions. However, the spatial resolution, field of view, scan time and number of gradient direction selected for this study are commonly used for FA-study and tractography purposes.