Abstract
The amounts of muscle and fat in a person's body, known as body composition, are correlated with cancer risks, cancer survival, and cardiovascular risk. The current gold standard for measuring body composition requires time-consuming manual segmentation of CT images by an expert reader. In this work, we describe a two-step process to fully automate the analysis of CT body composition using a DenseNet to select the CT slice and U-Net to perform segmentation. We train and test our methods on independent cohorts. Our results show Dice scores (0.95-0.98) and correlation coefficients (R=0.99) that are favorable compared to human readers. These results suggest that fully automated body composition analysis is feasible, which could enable both clinical use and large-scale population studies.
Introduction
"Body composition"—the specific ratio and distribution of muscle and fat—is a powerful predictor of cancer survival, surgical outcomes, and cardiovascular risk. The clinical gold standard for measuring this involves analyzing a single CT cross-section at the level of the third lumbar vertebra (L3).
However, finding the exact L3 slice and manually segmenting the different tissue compartments (muscle, visceral fat, and subcutaneous fat) is a significant bottleneck. It typically takes an expert analyst about 20 minutes per scan. This manual overhead makes it nearly impossible to conduct population-scale studies involving thousands of patients. This research presents a "hands-off" system designed to perform this task with the speed of a computer and the precision of a radiologist.
Methods
The proposed solution utilizes a two-stage "detect then segment" pipeline.
Automated Slice Selection
Instead of a simple classifier, the system uses a regression-based approach. A DenseNet (or ResNeXt) is trained to analyze individual 2D slices and predict their distance (offset) from the L3 landmark. A sigmoid function is used to focus the model’s attention on the region immediately surrounding L3. This regression approach allows the model to learn the spatial context of the spine, making slice selection more robust than simple identification.
Tissue Segmentation
Once the L3 slice is identified, it is processed by a U-Net architecture. To address the class imbalance between background and specific tissue types, a soft Dice maximization loss function was implemented. This ensures the model prioritizes the overlap between predicted and actual tissue boundaries.
Datasets
The study utilized two distinct cohorts to ensure generalizability:
- Dataset A (Training/Validation): 595 CT scans from pancreatic cancer patients, manually segmented using Slice-O-Matic software.
- Dataset B (Independent Testing): 534 CT scans from lymphoma patients at Massachusetts General Hospital, segmented using Osirix. This independent set tests the model's ability to generalize across different diseases and software tools.
Key Results
The automated pipeline demonstrates performance comparable to human experts across several metrics:
- Accuracy: The segmentation model achieved Dice scores between 0.95 and 0.98.
- Correlation: The correlation coefficient (R) between automated and manual measurements was 0.99, indicating near-perfect agreement.
- Throughput: The entire automated process takes approximately 1 second on a modern GPU, representing a 1,200x increase in speed compared to manual analysis.
- Localization: The L3 slice was identified with a mean error of only 9.4 mm, well within the height of a single vertebra.
Discussion and Conclusion
A key insight from this work is that regression-based slice selection outperforms classification by leveraging spatial context. Furthermore, the model's success on an independent lymphoma dataset after being trained on pancreatic cancer data suggests it learns fundamental anatomical features rather than disease-specific markers.
While the L3 slice is the current clinical standard, it remains a proxy for the whole body. Future work may explore analyzing the distribution of fat and muscle across the entire abdominal volume. Additionally, while robust, the model can still face challenges with extreme body types or unusual anatomical variations. Moving forward, integrating this pipeline into real-time clinical workflows could provide radiologists with automatic body composition metrics immediately following a scan.
Summary updated