Related Experiment Video
Updated: Jul 1, 2026
13:35
Segmentation and Linear Measurement for Body Composition Analysis using Slice-O-Matic and Horos
Published on: March 21, 2021
Multimodal Large Language Model for Zero-Shot L3 Body Composition Segmentation on CT: Improved Accuracy via Automated
Haruto Sugawara1, Akiyo Takada2,3,4, Shimpei Kato5
1Department of Medical Imaging, The Ottawa Hospital, University of Ottawa, 501 Smyth Road, Ottawa, ON, K1H 8L6, Canada. capy.sugawara@gmail.com.
Journal of Imaging Informatics in Medicine
|June 30, 2026
Summary
Multimodal large language models can segment L3 body composition on CT scans. Automated candidate selection improved accuracy for skeletal muscle and subcutaneous adipose tissue, but not visceral adipose tissue.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Accurate body composition analysis from CT scans is crucial for clinical assessments.
- Current segmentation methods can be labor-intensive and time-consuming.
Purpose of the Study:
- To evaluate zero-shot L3 body composition segmentation using a multimodal large language model (MLLM).
- To assess if automated candidate selection enhances segmentation accuracy.
Main Methods:
- Retrospective analysis of the TCIA Colorectal-Liver-Metastases CT dataset.
- Utilized gemini-3-pro-image-preview for candidate mask generation and gemini-3-pro-preview for selection.
- Compared model segmentation with radiologist-defined reference masks using Dice Similarity Coefficient (DSC).
Main Results:
- Automated candidate selection improved DSC for skeletal muscle (SM) and subcutaneous adipose tissue (SAT).
- Segmentation accuracy for visceral adipose tissue (VAT) did not significantly improve with automated selection.
- Performance, particularly for VAT, remained below interobserver agreement between radiologists.
Conclusions:
- Zero-shot L3 body composition segmentation using a general-purpose MLLM is feasible.
- Automated candidate selection offers improvements for SM and SAT segmentation accuracy.
- Further advancements are needed to match interobserver performance, especially for VAT segmentation.