Automated segmentation of human cervical-supraclavicular adipose tissue in magnetic resonance images

Abstract

Human brown adipose tissue (BAT), with a major site in the cervical-supraclavicular depot, is a promising anti-obesity target. This work presents an automated method for segmenting cervical-supraclavicular adipose tissue for enabling time-efficient and objective measurements in large cohort research studies of BAT. Fat fraction (FF) and R2∗ maps were reconstructed from water-fat magnetic resonance imaging (MRI) of 25 subjects. A multi-atlas approach, based on atlases from nine subjects, was chosen as automated segmentation strategy. A semi-automated reference method was used to validate the automated method in the remaining subjects. Automated segmentations were obtained from a pipeline of preprocessing, affine registration, elastic registration and postprocessing. The automated method was validated with respect to segmentation overlap (Dice similarity coefficient, Dice) and estimations of FF, R2∗ and segmented volume. Bias in measurement results was also evaluated. Segmentation overlaps of Dice = 0.93 ± 0.03 (mean ± standard deviation) and correlation coefficients of r > 0.99 (P < 0.0001) in FF, R2∗ and volume estimates, between the methods, were observed. Dice and BMI were positively correlated (r = 0.54, P = 0.03) but no other significant bias was obtained (P ≥ 0.07). The automated method compared well with the reference method and can therefore be suitable for time-efficient and objective measurements in large cohort research studies of BAT.