Whole-body Composition Profiling Using a Deep Learning Algorithm: Influence of Different Acquisition Parameters on...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate body composition profiling is crucial for assessing health and disease.
• Manual segmentation of body compartments in whole-body MRI (wbMRI) is time-consuming and labor-intensive.
• Automated segmentation methods are needed to improve efficiency and reproducibility.

Purpose of the Study:

• To develop and validate a deep learning algorithm for automated segmentation of body compartments in wbMRI.
• To assess the algorithm's performance across different acquisition parameters and scanners.
• To investigate the robustness and test-retest reliability of the automated segmentation.

Main Methods:

• A U-net convolutional neural network was designed for segmenting subcutaneous adipose tissue (SCAT), visceral adipose tissue (VAT), and total muscle mass (TMM).
• Twenty clinical wbMRI scans were manually segmented for training, validation, and testing.
• Performance was evaluated using the Sorensen-Dice coefficient (DSC) on diverse datasets and compared with manual segmentations.

Main Results:

• The algorithm achieved high performance, with DSC scores ranging from 0.93 for SCAT to 0.77 for VAT on the test dataset.
• Similar performance was observed across different scanners and protocols, with DSCs for VAT between 0.69-0.72 and SCAT between 0.83-0.91.
• No significant differences in body composition profiling were found for repetitive scans or variations in protocol parameters (P = 0.88-1).

Conclusions:

• A deep learning-based algorithm enables automated body composition profiling from wbMRI.
• The developed algorithm demonstrates robust and reproducible segmentation performance, comparable to manual expert segmentation.
• The method is reliable across a range of different acquisition parameters, facilitating widespread clinical application.