Subject-level segmentation accuracy weights for volumetric studies involving label fusion

1846 

Abstract Submission 

Christina Chen¹, Sandhitsu Das¹, M Tisdall¹, Fengling Hu¹, Andrew Chen², Paul Yushkevich¹, David Wolk¹, Russell Shinohara¹

¹University of Pennsylvania, Philadelphia, PA, ²Medical University of South Carolina, Charleston, SC

Christina Chen  
University of Pennsylvania
Philadelphia, PA

Sandhitsu Das  
University of Pennsylvania
Philadelphia, PA

M Tisdall  
University of Pennsylvania
Philadelphia, PA

Fengling Hu  
University of Pennsylvania
Philadelphia, PA

Andrew Chen  
Medical University of South Carolina
Charleston, SC

Paul Yushkevich  
University of Pennsylvania
Philadelphia, PA

David Wolk  
University of Pennsylvania
Philadelphia, PA

Russell Shinohara  
University of Pennsylvania
Philadelphia, PA

Region-specific changes in brain volume accompany many neurological diseases. Delineating these changes can link structural and clinical findings to illuminate disease biology and facilitate clinical management. For example, Alzheimer's disease (AD) is a neurodegenerative disorder characterized by neuropathological deposits that concentrate in regions such as the hippocampus [1, 6]. MRI volumetric studies have augmented these findings by establishing correlations between the magnitude of hippocampal atrophy and the severity of clinical disease [3].
Deriving hippocampal volumes from a neuroimaging study entails segmenting the hippocampus in each image. Manual segmentation is cumbersome and prone to errors. Automated atlas-based segmentation borrows information from expert-labeled images (atlases). Multi-atlas label fusion refines single-atlas segmentation by aggregating labels from multiple atlases.
One way of testing for group differences in hippocampal volume involves estimating the hippocampal volume for each subject and then regressing these estimates onto the group labels. However, this approach equalizes the contributions of all study subjects. Retaining poorly segmented images can attenuate power or induce bias, but discarding them altogether can potentially deplete the sample size. Ideally, we would like to weight each volume estimate according to its precision. We propose a novel, statistically motivated method of deriving subject-level precision weights by quantifying how resampling different sets of atlases perturbs each subject's volume estimate.

We used T1-weighted MRIs from the ADNI. In our first experiment, we compared 200 images from controls (CN) to 200 images from AD subjects. In our second experiment, we compared 150 images from amyloid-negative controls to 175 images from subjects with amyloid-positive early mild cognitive impairment (EMCI).
We performed neck-trimming, N4 bias correction, and skull-stripping on each image. For multi-atlas segmentation, we used the 35 OASIS [5] atlases provided for [4]. We fused the labels from the registered atlases via joint label fusion (JLF) [7].
For each image, we obtained 100 bootstrapped atlas collections by sampling from our 35 atlases, derived the hippocampal volume from the JLF segmentation produced by each atlas collection, computed the variance across these 100 volume estimates, and computed the inverse of this variance to weight the hippocampal volume estimate for each subject.
We validated our weights by assessing how incorporating them affects the type I error rate and power in detecting a disease status effect on hippocampal volume. To derive a p-value for this effect, we performed the weighted version of the Freedman-Lane permutation test [2]. In all of our experiments, we regressed the subjects' hippocampal volume estimates onto their disease status (CN vs. AD or CN vs. EMCI), ages, and intracranial volumes.

Our weighting and permutation procedure controls the type I error rate. Moreover, it confers significantly increased power compared to OLS. For example, for 160 samples, weighting improves the power to detect a difference in mean hippocampal volume between controls and AD subjects from 0.592 to 0.724 (Fig. 1). The power gain for differentiating between amyloid-negative controls and amyloid-positive EMCI subjects is even more impressive. For example, for 120 samples, weighting improves the power from 0.112 to 0.615 (Fig. 2).

We proposed a novel way of deriving subject-level weights that quantify the variability in volume estimates across joint segmentations produced by different atlas collections. We also demonstrated on real data that incorporating these weights significantly improves power for detecting a mean group difference in hippocampal volume. Thus, our method provides definite guidance on how to leverage the uncertainty information extracted from the segmentation procedure to facilitate testing hypotheses motivated by biological questions.

Methods Development ¹

Segmentation and Parcellation ²

Data analysis

MRI

Segmentation

Statistical Methods

STRUCTURAL MRI

Other - Measurement error