TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anatomy

Background:

• Accurate segmentation of anatomic structures in CT images is crucial for various clinical applications.
• Existing segmentation models often lack robustness across diverse datasets and a wide range of anatomic structures.

Purpose of the Study:

• To develop and validate a deep learning segmentation model for automatic and robust segmentation of major anatomic structures on body CT images.
• To create a comprehensive tool for applications like organ volumetry, disease characterization, and therapy planning.

Main Methods:

• A retrospective study utilized 1204 CT examinations to train an nnU-Net segmentation algorithm on 104 distinct anatomic structures.
• The model was evaluated using Dice similarity coefficients on a large, real-world dataset.
• A separate dataset of 4004 CT examinations was used to investigate age-dependent changes.

Main Results:

• The deep learning model achieved a high Dice score of 0.943 on the test set, demonstrating robust performance on diverse clinical data.
• The model significantly outperformed a publicly available segmentation model (0.932 vs 0.871).
• Significant correlations were found between age and organ volume/attenuation, providing insights into aging processes.

Conclusions:

• The developed model provides robust and accurate segmentation of 104 anatomic structures in CT images.
• The annotated dataset and toolkit are publicly available, facilitating further research and clinical implementation.
• This deep learning approach advances automated medical image analysis.