Source-linked AI summary
Fully-automated Body Composition Analysis in Routine CT Imaging Using 3D Semantic Segmentation Convolutional Neural Networks
Sven Koitka, Lennard Kroll, Eugen Malamutmann, Arzu Oezcelik, Felix Nensa
TL;DR
Routine abdominal CT contains clinically valuable body-composition information that is difficult to extract manually at scale. This study develops a fully automated 3D system combining multi-resolution U-Net segmentation with HU-based tissue subclassification, achieving strong segmentation and tissue-volume agreement. The system supports volumetric biomarkers across the abdomen rather than only at the L3 reference level.
Problem
Manual body-composition assessment is labor-intensive, while prior deep-learning approaches focused mainly on single L3 CT slices.
Method
The system segments five semantic abdominal CT regions with multi-resolution U-Net 3D models, then subclassifies adipose tissue and muscle using HU thresholds.
Results
0.9553 mean Sørensen Dice and intra-class correlation coefficients above 0.99 were achieved for semantic segmentation and subclassified tissue volumes, respectively.
Takeaways & Limitations
Fully automated volumetric analysis can provide stable body-composition biomarkers across the whole abdomen from routine CT imaging.
Takeaways & Limitations
The current system is limited to abdominal examinations, with future work planned for other body regions and thinner slices.
Abstract
from arXiv · showhide
Body tissue composition is a long-known biomarker with high diagnostic and prognostic value in cardiovascular, oncological and orthopaedic diseases, but also in rehabilitation medicine or drug dosage. In this study, the aim was to develop a fully automated, reproducible and quantitative 3D volumetry of body tissue composition from standard CT examinations of the abdomen in order to be able to offer such valuable biomarkers as part of routine clinical imaging. Therefore an in-house dataset of 40 CTs for training and 10 CTs for testing were fully annotated on every fifth axial slice with five different semantic body regions: abdominal cavity, bones, muscle, subcutaneous tissue, and thoracic cavity. Multi-resolution U-Net 3D neural networks were employed for segmenting these body regions, followed by subclassifying adipose tissue and muscle using known hounsfield unit limits. The Sørensen Dice scores averaged over all semantic regions was 0.9553 and the intra-class correlation coefficients for subclassified tissues were above 0.99. Our results show that fully-automated body composition analysis on routine CT imaging can provide stable biomarkers across the whole abdomen and not just on L3 slices, which is historically the reference location for analysing body composition in the clinical routine.
Keypoints
The study enables fully automated body composition analysis on routine abdominal CT scans, with strong semantic-region segmentation and tissue-volume agreement.
- Keypoints: Fully automated body composition analysis was performed on routine abdominal CT scans.
- Keypoints: 0.9553 mean Sørensen Dice score was achieved for semantic body-region segmentation.The best models segmented the abdominal cavity, bones, muscle, subcutaneous tissue, and thoracic cavity.
- Keypoints: Intra-class correlation coefficients for subclassified tissue volumes exceeded 0.99.
1 Introduction
Body tissue composition is a valuable biomarker, but manual routine extraction is labor-intensive. The study therefore targets automated, reproducible 3D volumetry from standard abdominal CT examinations.
- 1 Introduction: Manual tissue-composition assessment in routine imaging requires extraordinary human labor despite its diagnostic and prognostic value.
- 1 Introduction: The study aims to provide fully automated, reproducible, quantitative 3D body-composition volumetry from standard abdominal CT examinations.
2 Materials and Methods
The study combines 3D semantic segmentation with HU-based tissue subclassification using annotated abdominal CT volumes. It trains and evaluates U-Net variants on five semantic body regions and generates SAT, VAT, and muscle volumes.
- 2 Materials and Methods: 40 CT scans were used for training and 10 for testing, with every fifth axial slice annotated across six labels including five body regions.
- 2 Materials and Methods: Multi-resolution U-Net 3D and U-Net 3D architectures were trained for volumetric semantic segmentation.
- 2 Materials and Methods: The networks used augmented, cropped volumetric inputs and sliding-window inference with weighted aggregation of overlapping predictions.
- 2 Materials and Methods: HU preprocessing mapped CT intensities into normalized network inputs, with transfer-function choice strongly affecting segmentation performance.
- 2 Materials and Methods: Training combined softmax cross-entropy and generalized Sørensen Dice losses, excluding background from the Dice term.
- 2 Materials and Methods: Semantic regions were converted into SAT, VAT, and muscle volumes using HU thresholds and region membership.Muscle used -29 to 150 HU, while adipose tissue used -190 to -30 HU; adipose voxels in abdominal versus subcutaneous regions were classified as VAT versus SAT.
3 Results
The evaluated segmentation models achieved strong semantic body-region performance, while HU-window selection materially affected Dice scores. The resulting tissue quantification showed very high agreement and was designed for routine clinical reporting.
- 3.1 Model Evaluation: Ribs remained especially problematic because partial-volume effects made region assignment difficult even for human readers in 5mm CTs.Most exemplary slices nevertheless showed almost perfect segmentation boundaries.
- 3.2 Ablation Study: Over 0.95 average Dice was achieved on both cross-validation and test sets when multiple HU windows were combined as separate input channels.Increasing the HU intensity range consistently improved Dice scores; the abdominal window from -150 to 250 produced the lowest averages, 0.829 for cross-validation and 0.875 for testing.
- 3.3 Tissue Quantification Report: 0.999, 0.998, and 0.991 intraclass correlation coefficients were obtained for SAT, VAT, and muscle volumes, respectively.The system generated numeric per-slice values and a visual PDF report that was encapsulated into DICOM and sent to PACS.
4 Discussion
The study develops automated 3D body-composition volumetry for routine abdominal CT, achieving strong segmentation and tissue-volume agreement while exposing artifact and scope limitations.
- System output: The final report visualizes subcutaneous adipose tissue, visceral adipose tissue, and muscle volumes using color-coded tissue maps and per-slice volume summaries.SAT is red, VAT green, and muscle yellow in the visual report.
- Results: A mean Sørensen Dice coefficient of 0.9553 was achieved for automated segmentation of five abdominal semantic regions, with derived tissue volumetry exceeding 0.99 intraclass correlation.The segmented regions were abdominal cavity, bones, muscle, subcutaneous tissue, and thoracic cavity.
- Motivation: Manual single-slice measurements are subject to intra-patient and inter- and intra-rater variability, motivating automated analysis across more than one CT slice.The study frames automated volumetry as a way to provide quantitative biomarkers within routine clinical imaging.
- Contribution: The approach analyzes annotated data across the whole abdomen rather than relying only on historically used L3 slices, supporting volumetric body-composition analysis.Adipose tissue was subclassified from semantic regions using known Hounsfield-unit thresholds.
- Limitations: The dataset’s 5 mm slice thickness introduces partial-volume effects, while future work targets thinner slices and extension beyond abdominal regions toward whole-body analysis.The authors note that 5 mm reconstruction is common in routine clinical CT.
- Limitations: Fat within organs is currently counted as visceral adipose tissue when it lies inside the abdominal cavity, although it should be excluded from final VAT statistics.Multi-organ segmentation could support postprocessing by masking organs.
- Limitations: Beam-hardening artifacts can impair segmentation or invalidate HU-based muscle and fat identification, even when semantic-region boundaries remain accurate.The authors plan automatic artifact detection and handling in a future version.
5 Conclusion
The study presents a fully automated deep learning system for volumetric tissue classification in abdominal CT to extract robust biomarkers.
- The system enables fully automated volumetric tissue classification from abdominal CT examinations for extracting robust biomarkers.