Dynamic Fuzzy-Gaussian Modeling (DynFGM): A Kurtosis-Adaptive Unsupervised Framework for Automated Adipose Tissue...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate quantification of abdominal adipose tissue using MRI is crucial for metabolic risk assessment.
• Current methods like manual segmentation are labor-intensive, and deep learning models require extensive labeled data.
• There is a need for automated, data-efficient frameworks for adipose tissue segmentation.

Purpose of the Study:

• To introduce Dynamic Fuzzy-Gaussian Modeling (DynFGM), a fully automated, unsupervised framework for abdominal adipose tissue segmentation.
• To develop and validate DynFGM without requiring training data, expert annotations, or anatomical priors.
• To provide a mathematically interpretable alternative to deep learning models for adipose tissue phenotyping.

Main Methods:

• DynFGM utilizes image intensity kurtosis to dynamically adapt its complexity for each MRI slice.
• A fuzzy C-means (FCM) algorithm initializes a Gaussian mixture model (GMM) for segmentation.
• A radial distance transform differentiates subcutaneous (SAT) from visceral adipose tissue (VAT).

Main Results:

• DynFGM achieved high spatial agreement (mean DSC: 0.94) and volumetric reliability (ICC: 0.82-0.97) compared to expert annotations.
• The framework significantly reduced mean absolute volumetric error by 92.6% compared to standard FCM.
• Operational stability was demonstrated on a large cohort (n=756) with a low technical failure rate (3.0%) and fast computational throughput (13.6 s/participant).

Conclusions:

• DynFGM offers an interpretable and data-efficient unsupervised approach for abdominal adipose tissue phenotyping.
• It serves as a scalable tool for population-level research and an automated labeling solution where labeled data are limited.
• This framework bridges the gap between manual segmentation and supervised deep learning, facilitating future model development.