Automated Abdominal Segmentation of CT Scans for Body Composition Analysis Using Deep Learning

Visceral adipose tissue (VAT), which is linked with the metabolic consequences of obesity, is usually characterized by measuring VAT area at the L4-L5 vertebral interspace. However, the location of the slice with the strongest relation to VAT volume is not established. We sought to investigate the relations between cross-sectional VAT areas at different anatomic locations and VAT volume in a large, diverse sample of healthy subjects. VAT volume was derived from slice areas taken at 5-cm intervals from magnetic resonance images in 121 healthy men [x +/- SD age: 41.9 +/- 15.8 y; body mass index (BMI; in kg/m(2)): 26.0 +/- 3.2; VAT: 2.7 +/- 1.8 L] and 198 healthy women (age: 48.1 +/- 18.7 y; BMI: 27.0 +/- 5.4; VAT: 1.7 +/- 1.2 L). Regression models were developed to identify the best single slice for estimating VAT volume. The VAT area 10 cm above L4-L5 (A(+10)) in men (R(2) = 0.932, P < 0.001) and 5 cm above L4-L5 (A(+5)) in women (R(2) = 0.945, P < 0.001) had the highest correlation with abdominal VAT. R(2) increased by only 3.8% in men and 0.5% in women with adjustment for age, race, scanning position, BMI, and waist circumference. Studies using A(+10) in men and A(+5) in women will require 14% and 9% fewer subjects, respectively, than those using slices at L4-L5 and will have equivalent power. Measurement of slice areas at A(+10) in men and A(+5) in women provides greater power for the detection of VAT volume differences than does measurement at L4-L5.

Obesity, defined by body mass index (BMI), appears to have a paradoxical protective effect in several chronic diseases.