Development of automated segmentation of visceral adipose tissue in computed tomography

Imaging modalities such as computed tomography (CT) or magnetic resonance imaging have been used to measure adiposity. However, manual segmentation of visceral adipose tissue (VAT) in the entire abdomen is laborious and time-consuming. We aimed to develop a new method for accurate visceral fat segmentation by automatically dividing the three anatomical compartments of the lung, soft tissue, and post-vertebral spaces.

To automatically separate visceral fat, a three-step process was performed that sequentially divided tissues and regions in a three-dimensional CT image. Manual segmentation was performed in 99 individuals who underwent 18-fluoro-2-deoxyglucosepositron emission tomography/CT for cancer screening between January 2010 and December 2018 to validate the automated segmentation. The similarity index and Pearson's correlation analysis were performed to compare automated segmentation with manual segmentation. Clinical data, such as weight, height, and glucose and insulin levels, were measured. Pearson's correlation analysis was performed to investigate the association between the two methods.

VAT volume of automated segmentation (3,594.6 ± 1,776.5 cm³) strongly correlated with that of manual segmentation (3,375.7 ± 1567.5 cm³) (r = 0.9676, p < 0.0001). The similarity index positively correlated with the VAT volume (r = 0.6396, p < 0.0001) and negatively correlated with the mean Hounsfield units (HU) (r = -0.4328, p < 0.0001). Bland-Altman plots are presented with 5.1 % for VAT volume and 7.1 % for mean HU were outside 1.96 standard deviation from the mean value.

We developed an automated segmentation method for VAT in the entire abdomen. This automated segmentation method is feasible for measuring the VAT volume and VAT HU. This method could be employed in daily clinical practice to provide more detailed information about VAT.