Within-subject template estimation for unbiased longitudinal image analysis

Abstract

Longitudinal image analysis has become increasingly important in clinical studies of normal aging and neurodegenerative disorders. Furthermore, there is a growing appreciation of the potential utility of longitudinally acquired structural images and reliable image processing to evaluate disease modifying therapies. Challenges have been related to the variability that is inherent in the available cross-sectional processing tools, to the introduction of bias in longitudinal processing and to potential over-regularization. In this paper we introduce a novel longitudinal image processing framework, based on unbiased, robust, within-subject template creation, for automatic surface reconstruction and segmentation of brain MRI of arbitrarily many time points. We demonstrate that it is essential to treat all input images exactly the same as removing only interpolation asymmetries is not sufficient to remove processing bias. We successfully reduce variability and avoid over-regularization by initializing the processing in each time point with common information from the subject template. The presented results show a significant increase in precision and discrimination power while preserving the ability to detect large anatomical deviations; as such they hold great potential in clinical applications, e.g. allowing for smaller sample sizes or shorter trials to establish disease specific biomarkers or to quantify drug effects.

Copyright © 2012 Elsevier Inc. All rights reserved.

Figures

Fig. 1

Simplified diagram of the three steps involved in longitudinal processing, showing information flow at a single longitudinal run. Dashed line: information is used for initialization. Solid line: information is copied.

Fig. 2

Unbiased template estimation for a subject with neurodegenerative disease and significant atrophy: All time points are iteratively aligned to their median image with an inverse consistent robust registration method, resulting in a template image and simultaneously a co-registration of all time points.

Fig. 3

Comparison of mean and median template image for a series of 18 images (7 years) of a subject with neurodegenerative disorder (Huntington's disease). The difference image (top row) between median and mean reveals large differences in regions that change over time (e.g. ventricles, corpus callosum, eyes, neck, scalp). Below are close-ups of the mean image (left, softer edges) and the median image (right, crisper edges).

Fig. 4

Initializing time point 2 with results from time point 1 [BASE1] and vice-versa [BASE2] shows a bias in symmetrized percent change. Using our method [FS5.1] and passing time points in reverse order [FS5.1rev] does not show a processing bias. Significant differences from zero are denoted by a red plus: p<0.01 and red star: p<0.001 in Wilcoxon signed rank test. Error bars show a robust standard error where standard deviation is replaced by σ≈ 1.4826 median absolute deviation. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 5

Effect of simulated noise (σ= 1) on hippocampal volume measurements. The longitudinal processing is less affected.

Fig. 6

Simulated 2% atrophy in the left hippocampus. The longitudinal processing manages to detect the change more precisely and at the same time reduces the variability in the right hemisphere.

Fig. 7

Test–retest comparison of independent [CROSS] versus longitudinal [LONG] processing on TT-115 data (left hemisphere).Subcortical (left), cortical gray matter (middle), and white matter segmentations (right). The mean absolute volume difference (as percent of the average volume) is shown with standard error. [LONG] significantly reduces variability. Red dot: p<0.05, red plus: p<0.01, red star: p<0.001. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 8

Test–retest comparison of [CROSS] versus [LONG] on TT-14 data (subcortical volumes, left hemisphere). See also Fig. 7 for description of symbols. Additionally here reliability improvements of using a common voxel space (Long 5.1b) over previous method (Long 5.0) can be seen.

Fig. 9

Left: Average absolute symmetrized percent thickness change at each vertex for TT-115 using [CROSS]. Some regions (yellow) show 6% ASPC and above. Middle: Comparison: ([CROSS]-[LONG]) of average absolute symmetrized percent thickness change at each vertex for TT-115. Blue: [LONG] has larger variability, red/yellow [CROSS] has larger variability. [LONG] improves reliability in most regions, especially in the frontal and lateral cortex (yellow: more than 2% reduction of variability, frontal and lateral even more than 4%). Blue and red regions are mainly noise. Right: corresponding significance map, false discovery rate corrected at 0.05. [LONG] improves reliability significantly in most regions.

Fig. 10

Percent of subjects needed in [LONG] with respect to [CROSS] to achieve same power at same p to detect same effect size. In most regions less than 40% of the subjects are needed. Equivalently this figure shows the reduction in necessary time points when keeping the number of subjects and within-subject variance of time points the same. Variance of measurements and correlation were estimated based on TT-14 using bootstrap with 1000 samples. Bars show median and error bars depict 1st and 3rd quartile.

Fig. 11

Percent volume change per year with respect to baseline of the OA-136 dataset (2 to 5 visits per subject) for both independent [CROSS] (top) and longitudinal [LONG] (bottom) processing. [LONG] shows greater power to distinguish the two groups and smaller error bars (higher precision).

Fig. 12

Symmetric percent volume change per year of several subcortical structures. Left: [CROSS] almost no significant differences due to high variability (small group sizes). Middle: [LONG] significant differences between pre-symptomatic (PHD far from onset) and controls and between PHD far and PHD near (left caudate). Right: Volume means (ICV normalized) at baseline (tp1). While baseline volume distinguishes groups in several cases, the significant difference between controls and PHDfar in atrophy rate in the putamen cannot be detected in the baseline volumes.

Fig. A.13

Votes that need to agree on a different label to convince a time point to swap at σ= 3 for a given intensity difference.