Sparse unbiased analysis of anatomical variance in longitudinal imaging

Abstract

We present a new algorithm for reliable, unbiased, multivariate longitudinal analysis of cortical and white matter atrophy rates with penalized statistical methods. The pipeline uses a step-wise approach to transform and personalize template information first to a single-subject template (SST) and then to the individual's time series data. The first stream of information flows from group template to the SST; the second flows from the SST to the individual time-points and provides unbiased, prior-based segmentation and measurement of cortical thickness. MRI-bias correction, consistent longitudinal segmentation, cortical parcellation and cortical thickness estimation are all based on strong use of the subject-specific priors built from initial diffeomorphic mapping between the SST and optimal group template. We evaluate our approach with both test-retest data and with application to a driving biological problem. We use test-retest data to show that this approach produces (a) zero change when the retest data contains the same image content as the test data and (b) produces normally distributed, low variance estimates of thickness change centered at zero when test-retest data is collected near in time to test data. We also show that our approach--when combined with sparse canonical correlation analysis--reveals plausible, significant, annualized decline in cortical thickness and white matter volume when contrasting frontotemporal dementia and normal aging.

Fig. 1

The SUAAV algorithm’s pipeline requires a template that contains cortical priors and cortical labels that are used in a parcellation scheme. These priors are mapped first to a single-subject template (SST) and used to initialize segmentation of the SST. The resulting probability maps are then deformed from SST space to individual space to initialize a prior-constrained segmentation of (in this case) a baseline and follow-up image collected at a one year interval. Thickness maps and cortical parcellation are then computed for each time point image. While parcellation is not explicitly required by SUAAV, we use parcellation, here, to evaluate the stability of the pipeline in test-retest data and to verify the findings provided by the voxel-wise analysis. Red arrows point to a region associated—in the group analysis—with significant atrophy and black arrows point to the corresponding region in the parcellation.

Fig. 2

The SCCAN component of SUAAV is illustrated. The template and the masks for each “view” (Jacobian and thickness change) are at left. The top row shows a selection of individual Jacobian change images in the template space, after smoothing. The same subject’s thickness change images, after smoothing, are in the bottom row. A slice from the SCCAN weight maps for each view is shown at right.

Fig. 3

The SUAAV method provides a restricted set of regions over which to perform statistical testing. Results are FDR corrected where we accept q-value (corrected p-value) < 0.05 as significant with cluster size > 100 mm³. Left frontal cortex atrophies at a greater rate (approximately 6–10 % per year) in FTD than controls. The template labeling identifies the cortical regions as middle frontal and inferior frontal gyrus. The white matter loss occurs bilaterally in the anterior frontal lobe. These findings are consistent with what is known about frontotemporal dementia.