Water-fat MRI, based on multi-echo acquisition, enables quantitative and simultaneous estimations of fat fraction (FF) and R₂*.¹ In imaging of human brown adipose tissue (BAT), FF and R₂* reflect potential differences in water, fat and iron content between BAT and white adipose tissue (WAT), helping separating and characterizing the two adipose tissue types.^2,3 The segmentation of BAT depots from surrounding tissue, such as WAT and muscle, generally requires time-consuming manual delineation. A reliable automated method for segmentation of BAT, applicable for subjects with a wide range of body composition, would facilitate image analysis. To evaluate the performance of a fully automated method (based on multi-atlas registration⁴), using a semi-automated method (based on manual delineation) as reference, for segmentation of human cervical-supraclavicular adipose tissue (regarded as suspected BAT, hereafter denoted sBAT).

Seventeen subjects with a wide range of body composition were examined in a clinical whole-body 1.5 T MR system (Philips Healthcare, Best, The Netherlands) using a 3D multi gradient echo sequence (FOV (RL×AP×FH) = 480×200×50 mm³, acq/rec voxel size = 1.0×1.0×2.0 mm³, 25 slices)⁵. Coregistered water images, fat images, FF maps and R₂* maps were obtained by 3D reconstruction using a previously described water-fat separation algorithm.⁶ The subjects were divided into two groups: a multi-atlas (MA) group (5 males, 4 females, age 23.1±10.4 years, BMI 26.1±5.7 kg/m²) and an evaluation (E) group (3 males, 5 females, 20.8±7.9 years, BMI 26.7±5.7 kg/m²).

Segmentation of sBAT: Initially, crude segmentation of anatomically defined sBAT volumes of interest (sBAT VOIs) was performed manually.⁵ The VOIs of the MA group were used for the automated segmentation and the VOIs of the E group were used for evaluation of the automated segmentation. Transfer of the crude sBAT VOI from each data set in the MA group to each data set in the E group was accomplished by image registration (see section Image registration below). After transferring the multi-atlas sBAT VOIs to an evaluation data set, the individual multi-atlas sBAT VOIs were combined to a total multi-atlas sBAT VOI through majority voting. Subsequently, final segmentation of the adipose tissue within the sBAT VOIs was performed automatically by range limits on FF (≥40%), followed by morphological 3D erosion and range limits on R₂* (≤50 s^-1).⁵ This final step resulted in the segmented sBAT VOIs: MA_i (from automated multi-atlas segmentation) and E_i (from semi-automated segmentation), for each member i of the E group.

Image registration: A two-step image registration method was performed using the Elastix software⁷, where a target (multi-atlas) volume was deformed to match a reference (evaluation) volume by optimization of a cost function. In the first step, the normalized correlation between the water fraction (WF=1-FF) maps, from which skin had been removed by erosion and background had been removed by thresholding, was optimized under an affine deformation. In the second step, a weighted sum of the normalized correlation between the WF maps and median filtered R₂* maps, from which skin had been removed by erosion, was optimized under a deformable deformation.

Evaluation: Comparison of FF and R₂* values, obtained from M_i and E_i, was performed by linear regression and t-tests. P values <0.05 were considered statistically significant. The dice coefficient was used to evaluate the overlap between MA_i and E_i.

FF and R₂* values obtained by the automated segmentation correlated with those obtained by the semi-automated segmentation (r_FF=0.998 and r_R2* =0.985 (Figure 1a-b)). The automated segmentation method showed a significant but marginally lower FF (-0.35±0.37 pp, P=0.032) and tendency of a higher R₂* (0.29±0.37 s^-1, P=0.062). The mean dice coefficient for the automated segmentation was 0.79±0.07.The strong correlation between FF and R₂* values obtained by the automated and semi-automated segmentation indicates a sufficient performance of the automated segmentation. However, there was a marginal underestimation of FF and a potential overestimation of R₂* suggesting a systematic error that warrants further investigation. The dice coefficients indicate moderate overlap between MA_i and E_i. This result was somewhat expected as this anatomical region is very flexible and sensitive to subject positioning and also due to the heterogeneity of the subjects in terms of body composition. Additionally, as the segmented sBAT VOIs were elongated in shape, small mismatches between M_i and E_i resulted in a relatively large reduction of the dice coefficients. Future work will include further optimization and evaluation of the automated segmentation method. The presented method shows promising results for automated segmentation of the human cervical-supraclavicular adipose tissue depot that allows time-efficient and objective measurements of BAT in large cohort research studies.