Automated Supraclavicular Brown Adipose Tissue Segmentation in Computed Tomography Using nnU-Net: Integration with...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Metabolic Diseases

Background:

• Brown adipose tissue (BAT) is crucial for energy expenditure and thermoregulation, making it a target for metabolic disease research.
• Manual segmentation of BAT in medical images is labor-intensive and prone to variability.

Purpose of the Study:

• To develop and evaluate an automated brown adipose tissue (BAT) segmentation method using the nnU-Net deep learning framework.
• To assess the performance of this automated method in a large cohort of lymphoma patients.

Main Methods:

• A 3D nnU-Net model was trained on 159 lymphoma patient CT scans with manual BAT annotations.
• An ensemble model was created and tested on an independent cohort of 30 patients, using DICE and Hausdorff Distance (HD) for evaluation.
• Standardized uptake values (SUVs) in BAT were analyzed in 7107 FDG PET/CT scans.

Main Results:

• The ensemble model achieved a high DICE score (0.780 ± 0.077) and acceptable HD (29.0 ± 14.6 mm), outperforming individual models.
• Automated segmentation revealed significant differences in BAT SUVs by sex, time of day, season, and age.
• Higher BAT SUVs were observed in women, morning scans, winter season, and younger individuals.

Conclusions:

• The automated BAT segmentation tool offers robust performance and reduces manual annotation effort.
• The tool's analysis of a large patient cohort validates known BAT SUV patterns.
• This method shows potential for widespread clinical and research applications in metabolic disease studies.