Gender-specific data-driven adiposity subtypes using deep-learning-based abdominal CT segmentation.

ABSTRACT

OBJECTIVE: The aim of this study was to quantify abdominal adiposity and generate data-driven adiposity subtypes with different diabetes risks. METHODS: A total of 3817 participants from the Pinggu Metabolic Disease Study were recruited. A deep-learning-based recognition model on abdominal computed tomography (CT) images (A-CT model) was developed and validated in 100 randomly selected cases. The volumes and proportions of subcutaneous fat, visceral fat, liver fat, and muscle fat were automatically recognized in all cases. K-means clustering was used to identify subgroups using the proportions of the four fat components. RESULTS: The Dice indices among the measurements assessed by the A-CT model and manual evaluation to detect liver fat, muscle fat, and subcutaneous fat areas were 0.96, 0.95, and 0.92, respectively. Three subtypes were generated separately in men and women visceral fat dominant type (VFD); subcutaneous fat dominant type (SFD); and intermuscular fat dominant type (MFD). Compared with the SFD group, the MFD group had similar diabetes risk, and the VFD group had a 60% higher diabetes risk when age and BMI were adjusted for in men. The adjusted odds ratio for diabetes was 1.92 (95% CI 1.32-2.78) in the MFD group and 6.14 (95% CI 4.18-9.03) in the VFD group in women. CONCLUSIONS: This study identified gender-specific abdominal adiposity subgroups, which may help clinicians to distinguish diabetes risk quickly and automatically.