Automated abdominal adipose tissue segmentation and volume quantification on longitudinal MRI using 3D convolutional...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Obesity research

Background:

• Increased subcutaneous and visceral adipose tissue (SAT/VAT) are linked to cardiometabolic disease risk.
• Accurate quantification of SAT/VAT is crucial for monitoring disease progression and treatment efficacy.
• Current segmentation methods can be time-consuming and labor-intensive.

Purpose of the Study:

• To develop and evaluate automated abdominal SAT/VAT segmentation using deep learning models.
• To assess the performance of attention-based competitive dense (ACD) 3D U-Net and 3D nnU-Net.
• To validate the models for longitudinal analysis in adults with overweight/obesity.

Main Methods:

• Utilized 920 adults with overweight/obesity undergoing two MRI scans.
• Trained and tested ACD 3D U-Net and 3D nnU-Net models on multi-contrast MRI data.
• Evaluated segmentation accuracy using Dice scores, false negatives/positives, and intraclass correlation coefficients for longitudinal volume agreement.

Main Results:

• Both ACD 3D U-Net and 3D nnU-Net achieved rapid segmentation (<4.8s/subject).
• High accuracy was demonstrated with median Dice scores ≥0.976 for VAT and ≥0.992 for SAT.
• Excellent longitudinal agreement for SAT/VAT volume was observed (ICC > 0.997).

Conclusions:

• ACD 3D U-Net and 3D nnU-Net provide automated, rapid, and accurate tools for quantifying abdominal SAT/VAT.
• These models are suitable for longitudinal analysis in adults with overweight/obesity.
• The developed methods can aid in the clinical assessment and management of cardiometabolic diseases.