Asymmetric bias in user guided segmentations of brain structures

Abstract

Brain morphometric studies often incorporate comparative hemispheric asymmetry analyses of segmented brain structures. In this work, we present evidence that common user guided structural segmentation techniques exhibit strong left-right asymmetric biases and thus fundamentally influence any left-right asymmetry analyses. In this study, MRI scans from ten pediatric subjects were employed for studying segmentations of amygdala, globus pallidus, putamen, caudate, and lateral ventricle. Additionally, two pediatric and three adult scans were used for studying hippocampus segmentation. Segmentations of the sub-cortical structures were performed by skilled raters using standard manual and semi-automated methods. The left-right mirrored versions of each image were included in the data and segmented in a random order to assess potential left-right asymmetric bias. Using shape analysis we further assessed whether the asymmetric bias is consistent across subjects and raters with the focus on the hippocampus. The user guided segmentation techniques on the sub-cortical structures exhibited left-right asymmetric volume bias with the hippocampus displaying the most significant asymmetry values (p≪0.01). The hippocampal shape analysis revealed the bias to be strongest on the lateral side of the body and medial side of the head and tail. The origin of this asymmetric bias is considered to be based in laterality of visual perception; therefore segmentations with any degree of user interaction contain an asymmetric bias. The aim of our study is to raise awareness in the neuroimaging community regarding the presence of the asymmetric bias and its influence on any left-right hemispheric analyses. We also recommend reexamining previous research results in the light of this new finding.

Low Plasma ApoE Levels Are Associated with Smaller Hippocampal Size in the Alzheimer's Disease Neuroimaging Initiative Cohort

“…While prior PiB studies in MCI and AD have not conclusively shown hemispheric asymmetry in amyloid deposition [44], PiB signal in the left, but not right, hemisphere in AD patients correlates with both cognitive performance and regional hypometabolism on FDG PET [45]. Alternatively, it has also been suggested that manual hippocampal volumetric analyses may be subject to asymmetric left-right biases [46]. This phenomenon could have theoretically impacted our results, even though our hippocampal indices were derived from automated segmentation, since the initial training set was developed through manual tracing.…”

Association of Depressive Symptoms with Hippocampal Volume in 1936 Adults

“…In addition, the study controlled for medication use, alcohol use, education, and other variables that might influence the relationship between depression and hippocampal volume. Another potential strength is the use of an automated method to derive hippocampal volumes that results in excellent reproducibility (Lucarelli et al, 2012;Nugent et al, 2012) and avoids a left-right bias that may be inherent with manual segmentation (Maltbie et al, 2012).…”

MR Imaging of Hippocampal Asymmetry at 3T in a Multiethnic, Population-Based Sample: Results from the Dallas Heart Study

“…The authors proposed "laterality of visual perception" as an inherent limitation in any study using user interaction in hippocampal segmentation and further suggested a reappraisal of prior work in light of this finding. 1 Thus, data previously reported by using manual segmentation have the potential for material bias.…”

“…1 At present, to our knowledge, there are no large studies with broad ethnic representation using high-resolution MR imaging, with consistent application of automated measures of the hippocampus to define the range of values of hippocampal asymmetry in the general population.…”