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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Metabolic Disease Research

Background:

• Abdominal adiposity is a key indicator of metabolic disease risk.
• Current methods for quantifying abdominal fat distribution are often labor-intensive and lack standardization.
• Identifying distinct adiposity subtypes could improve risk stratification for conditions like diabetes.

Purpose of the Study:

• To quantify abdominal adiposity using a novel deep-learning approach.
• To generate data-driven, gender-specific adiposity subtypes.
• To assess the differential diabetes risk associated with these identified subtypes.

Main Methods:

• Developed and validated a deep-learning model (A-CT model) for automated analysis of abdominal CT images.
• Quantified volumes and proportions of subcutaneous, visceral, liver, and muscle fat in 3817 participants.
• Employed K-means clustering to define adiposity subtypes based on fat component proportions.

Main Results:

• The A-CT model demonstrated high accuracy (Dice indices 0.92-0.96) compared to manual evaluation.
• Three distinct subtypes were identified in both men and women: visceral fat dominant (VFD), subcutaneous fat dominant (SFD), and intermuscular fat dominant (MFD).
• Compared to SFD, MFD showed similar diabetes risk, while VFD significantly increased diabetes risk (60% higher in men; OR 6.14 in women).

Conclusions:

• This study successfully generated gender-specific abdominal adiposity subgroups using AI-driven image analysis.
• These subtypes (VFD, SFD, MFD) are associated with distinct diabetes risks.
• The findings offer a potential tool for rapid, automated clinical assessment of diabetes risk based on abdominal fat distribution.