CAFT: a deep learning-based comprehensive abdominal fat analysis tool for large cohort studies.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Obesity Research

Background:

• Obesity is linked to sarcopenia and various comorbidities, including cancer, diabetes, hypertension, and stroke.
• Understanding fat distribution (subcutaneous adipose tissue [SAT] and visceral adipose tissue [VAT]) is key for metabolic health.
• Manual segmentation of fat depots is time-consuming and labor-intensive.

Purpose of the Study:

• To develop a deep learning (DL) pipeline for automated MRI-based fat quantification.
• To enable efficient analysis of SAT and VAT in large cohort studies.
• To provide a comprehensive data processing pipeline including augmentation, model training, visualization, and correction.

Main Methods:

• Utilized MRI data from 190 healthy older adults (mean age 67.85 years).
• Trained Residual Global Aggregation-based 3D U-Net (RGA-U-Net) and standard 3D U-Net models to segment SAT, VAT, superficial SAT (SSAT), and deep SAT (DSAT).
• Employed data augmentation to increase training datasets and evaluated accuracy using Dice and Hausdorff metrics.

Main Results:

• Achieved high segmentation accuracy: SSAT (0.92), DSAT (0.88), and VAT (0.9).
• Demonstrated strong correlation (R² > 0.99, p < 0.001) between automated and manual segmentations for all fat compartments.
• Showcased average Hausdorff distances below 5 mm and predicted volumes within ±1.96 SD via Bland-Altman analysis.

Conclusions:

• The DL-based tool accurately and reproducibly quantifies SSAT, DSAT, and VAT.
• Provides comprehensive fat compartment composition analysis and visualization in under 10 seconds.
• Offers a significant advancement for large-scale obesity and sarcopenia research.