Automated segmentation of five different body tissues on computed tomography using deep learning.

Area of Science:

• Medical Imaging Analysis
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accurate segmentation of body tissues in CT scans is crucial for quantitative analysis.
• Existing methods may lack efficiency or accuracy in segmenting multiple tissue types simultaneously.

Purpose of the Study:

• To develop and validate a novel computer tool for the automatic and simultaneous segmentation of five key body tissues: visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), intermuscular adipose tissue (IMAT), skeletal muscle (SM), and bone.
• To compare the performance of simultaneous versus separate segmentation of these tissues using deep learning models.

Main Methods:

• A dataset of 100 CT scans was utilized, with five body tissues manually annotated.
• Convolutional Neural Networks (CNNs), including UNet, R2Unet, and UNet++, were trained and validated using 10-fold cross-validation.
• A training-while-annotating strategy and 3D patch sampling were employed to optimize model development.

Main Results:

• Simultaneous segmentation of the five tissues yielded significantly higher Dice coefficients across all tissue types compared to separate segmentation (p < 0.05).
• Achieved Dice coefficients for simultaneous segmentation ranged from 0.826-0.840 for VAT, 0.901-0.908 for SAT, 0.574-0.611 for IMAT, 0.874-0.889 for SM, and 0.870-0.884 for bone.
• No significant performance differences were observed among the evaluated CNN models when segmenting tissues jointly.

Conclusions:

• Joint segmentation of visceral adipose tissue, subcutaneous adipose tissue, intermuscular adipose tissue, skeletal muscle, and bone on CT scans significantly outperforms separate segmentation.
• The developed computer tool demonstrates the efficacy of deep learning for simultaneous multi-tissue segmentation in radiological imaging.