In vivo analysis of hippocampal subfield atrophy in mild cognitive impairment via semi-automatic segmentation of T2-weighted MRI

Abstract

The measurement of hippocampal volumes using MRI is a useful in-vivo biomarker for detection and monitoring of early Alzheimer's disease (AD), including during the amnestic mild cognitive impairment (a-MCI) stage. The pathology underlying AD has regionally selective effects within the hippocampus. As such, we predict that hippocampal subfields are more sensitive in discriminating prodromal AD (i.e., a-MCI) from cognitively normal controls than whole hippocampal volumes, and attempt to demonstrate this using a semi-automatic method that can accurately segment hippocampal subfields. High-resolution coronal-oblique T2-weighted images of the hippocampal formation were acquired in 45 subjects (28 controls and 17 a-MCI (mean age: 69.5 ± 9.2; 70.2 ± 7.6)). CA1, CA2, CA3, and CA4/DG subfields, along with head and tail regions, were segmented using an automatic algorithm. CA1 and CA4/DG segmentations were manually edited. Whole hippocampal volumes were obtained from the subjects' T1-weighted anatomical images. Automatic segmentation produced significant group differences in the following subfields: CA1 (left: p = 0.001, right: p = 0.038), CA4/DG (left: p = 0.002, right: p = 0.043), head (left: p = 0.018, right: p = 0.002), and tail (left: p = 0.019). After manual correction, differences were increased in CA1 (left: p < 0.001, right: p = 0.002), and reduced in CA4/DG (left: p = 0.029, right: p = 0.221). Whole hippocampal volumes significantly differed bilaterally (left: p = 0.028, right: p = 0.009). This pattern of atrophy in a-MCI is consistent with the topography of AD pathology observed in postmortem studies, and corrected left CA1 provided stronger discrimination than whole hippocampal volume (p = 0.03). These results suggest that semi-automatic segmentation of hippocampal subfields is efficient and may provide additional sensitivity beyond whole hippocampal volumes.

Figure 1

Coronal slice of the hippocampal body in a T1 scan (left; 1×1 mm in-plane resolution) and a T2 scan (right; 0.2×0.2 mm in-plane resolution) obtained from a 3T scanner. Both images were collected from the same subject in one session. The dark band that is necessary to separate CA from DG is completely indiscernible in the T1 image, while it is clear in the T2 image.

Figure 2

Slice showing a section of the hippocampal body with a cyst between CA1 (red) and DG (blue).

Figure 3

Coronal section of the hippocampal body with automatic segmentation (center) and manually corrected segmentation (right). The two methods differ mainly in selecting the dark band boundary. Also included in this slice is a hippocampal cyst, which the automatic method incorrectly includes in CA1 and DG labels, while manual segmentation excludes it altogether.

Figure 4

A single hippocampus from two control subjects, and a representative image of the slices captured by the T2 sequence. Head slices are in blue, body slices in red, and tail slices in purple. The shape of the hippocampus and heuristic rules for segmentation determine how slices are assigned to a particular region. The hippocampal volume and ICV in each subject are nearly identical, but the distribution of volume is not even between regions, resulting in fewer body slices in the left subject. Thus the absolute volume of CA1 would appear much lower in this subject, though in reality they are likely quite similar. To account for this, volume is normalized by the number of slices in a region.

Figure 5

Figure 5a: Plots of normalized CA1 volume versus cohort. Volumes for the left hippocampus are in the left column of the figure. Automatically generated volumes are in the top row and manually corrected volumes in the bottom. In both cases, automatic segmentation detects statistically significant differences in volume by cohort, with manual correction displaying increased discrimination. Figure 5b: Plots of normalized CA4/DG volume versus cohort. Automatic segmentation (top row) demonstrates significant results bilaterally, while manual correction reveals no significant difference in the right side and only a subtle effect of cohort on the left.

Figure 5

Figure 5a: Plots of normalized CA1 volume versus cohort. Volumes for the left hippocampus are in the left column of the figure. Automatically generated volumes are in the top row and manually corrected volumes in the bottom. In both cases, automatic segmentation detects statistically significant differences in volume by cohort, with manual correction displaying increased discrimination. Figure 5b: Plots of normalized CA4/DG volume versus cohort. Automatic segmentation (top row) demonstrates significant results bilaterally, while manual correction reveals no significant difference in the right side and only a subtle effect of cohort on the left.

Figure 6

ROC curves and AUC values for left and right whole hippocampal volumes, CA1, and manually corrected CA1. Left manually corrected CA1 is a significantly better discriminant than whole hippocampal volume.