Comparison of Infant FreeSurfer segmentation and manual tracing in alcohol-exposed neonate brains

2014 

Abstract Submission 

Fleur Warton¹, Frances Robertson², R. Carter³, Lilitha Cengani¹, André van der Kouwe⁴, Lilla Zollei⁵, Neil Dodge⁶, Joseph Jacobson⁶, Sandra Jacobson⁶, Ernesta Meintjes¹

¹University of Cape Town, Cape Town, South Africa, ²Cape Universities Body Imaging Centre, University of Cape Town, Cape Town, South Africa, ³Columbia University, New York, NY, ⁴Athinoula A. Martinos Center for Biomedical Imaging, MGH, HMS, Boston, MA, ⁵Harvard University, Boston, MA, ⁶Wayne State University, Detroit, MI

Fleur Warton, PhD  
University of Cape Town
Cape Town, South Africa

Frances Robertson, Phd  
Cape Universities Body Imaging Centre, University of Cape Town
Cape Town, South Africa

R. Carter, MD  
Columbia University
New York, NY

Lilitha Cengani  
University of Cape Town
Cape Town, South Africa

André van der Kouwe  
Athinoula A. Martinos Center for Biomedical Imaging, MGH, HMS
Boston, MA

Massachusetts General Hospital Zollei, Dr.  
Harvard University
Boston, MA

Neil Dodge, PhD  
Wayne State University
Detroit, MI

Joseph Jacobson, PhD  
Wayne State University
Detroit, MI

Sandra Jacobson, PhD  
Wayne State University
Detroit, MI

Ernesta Meintjes, PhD  
University of Cape Town
Cape Town, South Africa

Segmentation of infant MRI brain images is challenging due to low signal-to-noise ratio, motion corruption, and maturation-related regional variation in tissue contrast and geometry¹. However, neonatal morphometric brain changes are associated with teratogenic exposures^2,3 and later neurodevelopment^4,5. Infant-specific tools are thus needed as manual segmentation is not feasible for large datasets. Infant FreeSurfer (IFS) is an automated segmentation pipeline for image processing in 0-2 year-olds⁶. Here we assess the reliability of IFS segmentation in comparison to manual segmentation of subcortical brain regions in neonates with prenatal alcohol exposure (PAE), and their relative sensitivity to detect PAE-related regional volume changes.

64 neonates (mean gestational age at scan: 41.4±2.3 wk) born to heavy-drinking or control women were scanned unsedated on a 3T Siemens Allegra MRI scanner with a custom-built neonatal head coil⁷. A motion-navigated multiecho gradient echo sequence was acquired twice with protocol parameters: FOV 144mm, 128 slices, TR 20ms, TE 1.46/ 3.14/ 4.82/ 6.5/ 8.18/ 9.86/ 11.54/ 13.22ms, 1mm³ isotropic resolution, flip angles 5° and 20° respectively. In FreeSurfer⁸, individual echoes from each acquisition were split, tissue parameters estimated and image volumes synthesised with a 24° flip angle. Volumes of bilateral caudate, putamen, pallidum, accumbens, thalamus, hippocampus, amygdala and cerebellar hemispheres, and midbrain, pons, medulla and vermis, were obtained from manual tracing in Freeview in FreeSurfer⁹ and automated segmentation in IFS⁶. Reliability across methods was assessed using intraclass correlation (ICC) estimates for consistency and absolute agreement with a single-rating two-way mixed effects model and Cronbach's alpha. Sensitivity to detect alcohol-related differences was assessed using a repeated measures general linear model (GLM) for each region with main effects of diagnostic group (27 unexposed, 23 heavily-exposed non-syndromal [HE], 14 pFAS/FAS; diagnosed in early childhood) and segmentation method, and group by method interactions. To examine sensitivity in detecting PAE effects, we performed ANOVA, with ANCOVA to control for potential confounding, on data obtained from each segmentation method separately to identify regions showing volumetric group differences.

Consistency between segmentation methods was moderate to excellent (> 0.5) in 60% of regions, but agreement poor (< 0.5) in 70% (Table 1). Cerebellar and brainstem regions generally showed acceptable reliability, while pallidum, accumbens and amygdala showed poor reliability bilaterally across all ICC measures. GLM analyses showed a significant main effect of segmentation method in all regions other than left hippocampus and medulla (Table 1), with manual segmentations smaller than IFS in all regions except cerebellum and brainstem. A main effect of diagnosis (p < 0.1) was observed in midbrain, medulla, and bilateral putamen and pallidum. Method*diagnosis interaction effects were seen in four regions. Fig. 1 shows regions where either manual or IFS volumes differed by group (p < 0.1). Post hoc pairwise comparisons showed smaller volumes in HE and/or pFAS/FAS groups for 7 manually traced and 9 IFS regions. After controlling for potential confounders, group differences remained in manually traced medulla and bilateral pallidum and in IFS-segmented bilateral caudate, left hippocampus, right amygdala, right accumbens, midbrain and medulla, with pFAS/FAS volumes smaller than unexposed in all but 1 region.

Our results show moderate region-dependent reliability between manual tracing and IFS segmentation. Both methods detected PAE-related volume reductions across basal ganglia, limbic and brainstem regions, but effects were more robust in IFS segmentations.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ²

Segmentation and Parcellation ¹

Neuroanatomy Other

Anatomical MRI

Data analysis

PEDIATRIC

Segmentation

STRUCTURAL MRI

Other - Prenatal alcohol exposure; neonate; manual segmentation