A Pre-Trained Model Customization Framework for Accelerated PET/MR Segmentation of Abdominal Fat in Obstructive Sleep...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate quantification of visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) is crucial for understanding cardiometabolic diseases like obstructive sleep apnea (OSA).
• Deep learning models accelerate repetitive tasks, but customization frameworks are needed for specific research applications.

Purpose of the Study:

• To validate a customization framework for developing automated VAT/SAT segmentation models using hybrid PET/MRI data.
• To assess the efficiency and accuracy of AI-driven segmentation compared to manual methods.

Main Methods:

• A UNet-ResNet50 model pre-trained on RadImageNet was iteratively trained on annotated PET/MRI scans.
• Model performance was evaluated against manual annotations using Dice similarity coefficients, segmentation time, and ICC for volumetric and metabolic agreement.

Main Results:

• Fully automated AI segmentation reduced annotation time from 121.8 min (manual) to 1.2 min per scan.
• High Dice similarity coefficients (0.98 for VAT/SAT masks) and excellent agreement (ICCs > 0.98) were achieved for volumetric and metabolic measures.
• Minimal bias was observed between AI-derived and manual measurements.

Conclusions:

• The developed AI pipeline enables efficient and accurate abdominal fat quantification using hybrid PET/MRI for simultaneous volumetric and metabolic analysis.
• This framework streamlines research workflows for obesity, OSA, and cardiometabolic disease studies by integrating multi-modal imaging and AI segmentation.
• Quantification of depot-specific adipose metrics can be facilitated, potentially influencing clinical outcomes.