Longitudinal variability analysis on low-dose abdominal CT with deep learning-based segmentation

Xin Yu, Yucheng Tang, Qi Yang, Ho Hin Lee, Riqiang Gao, Shunxing Bao, Ann Zenobia Moore, Luigi Ferrucci, Bennett A. Landman

TL;DR

Researchers developed an automated deep learning pipeline to segment 13 different abdominal tissues from low-dose 2D CT scans. By analyzing scans taken two years apart, they discovered that while muscle and fat measurements are highly reliable longitudinal markers, organ measurements are extremely sensitive to the specific axial slice location, posing a challenge for long-term clinical tracking.

What Problem Does This Paper Address?

Metabolic health is a cornerstone of modern medicine, linked to everything from heart disease to neurodegeneration. To study these links, researchers often look at "body composition"—the distribution of muscle, fat, and organs.

While 3D CT scans provide a full picture, they expose patients to significant radiation. Consequently, many longitudinal studies (studies that follow participants over years) use a single 2D "slice" of the abdomen to estimate body composition. However, a major question remains: How much of the change we see over time is actual biological change, and how much is just "noise" caused by slight differences in where the slice was taken or errors in the automated segmentation? This paper provides the first comprehensive look at this variability using automated deep learning tools.

Methods

The authors built a hybrid segmentation pipeline to handle 13 different tissue types:

  • Supervised Deep Learning: They used DeepLab-v3 (with a ResNet101 backbone) to segment large organs (liver, spleen, kidneys, stomach, aorta) and the abdominal wall. A standard U-Net architecture was employed specifically for muscle segmentation.
  • Unsupervised Clustering: For fat tissues—Subcutaneous Fat (SFT), Visceral Fat (VFT), and Retroperitoneal Fat (RFT)—and the overall body mask, they used a two-stage fuzzy c-means clustering method. This approach classifies pixels based on their Hounsfield Unit (HU) intensities without needing manual labels for every training image.
  • Statistical Validation: To measure longitudinal variability, they used the Intraclass Correlation Coefficient (ICC) (measuring agreement between scans) and the Coefficient of Variation (CV) (measuring the spread of data).

Datasets

The study utilized the Baltimore Longitudinal Study on Aging (BLSA) dataset:

  • Total Data: 1,816 2D CT slices from 1,469 subjects.
  • Training/Testing: Subsets of 1,175 and 300 subjects were used to train the multi-organ and muscle models, respectively.
  • Longitudinal Cohort: 300 subjects who had two scans taken exactly two years apart were selected to evaluate how measurements changed over time.

Key Insights and Ideas

  • The "Slice Position" Trap: The most striking finding was that organ measurements (like the stomach or spleen) had very low reliability (ICC < 0.5). The authors realized this wasn't necessarily due to biological changes, but because organs change shape rapidly along the vertical (z-axis) of the body. If the second scan is taken just a few millimeters higher or lower than the first, the organ's cross-sectional area changes drastically.
  • Muscle and Fat are Robust: Unlike organs, muscle and fat layers are relatively consistent across adjacent slices. This makes them excellent candidates for 2D longitudinal studies, as they are less sensitive to minor positioning errors.
  • Segmentation Stability: The deep learning models achieved high Dice scores (0.82 to 0.96), proving that the automated tools are consistent. The variability in the results comes from the underlying anatomy and imaging physics, not the AI's "uncertainty."

Key Results

  • High Reliability: Muscle, abdominal wall, fat, and body mask all showed high agreement over two years, with an average ICC ≥ 0.8.
  • Low Reliability: Organs showed poor agreement. The stomach was the most volatile, with an ICC of nearly 0.003 for area, indicating almost no correlation between the two-year scans due to its irregular shape and position.
  • Quantitative Accuracy: The muscle area showed a low Coefficient of Variation (~4%), while the stomach area CV was much higher (~29%).
  • Segmentation Performance: The pipeline successfully segmented 13 tissues with high accuracy, providing a robust tool for future large-scale epidemiological studies.

Limitations and Future Work

The primary limitation is the inherent "2D problem." Because 2D slices are so sensitive to vertical positioning, they may never be truly reliable for tracking organ changes unless the slice selection is perfectly standardized (e.g., tied to a specific vertebral level).

The authors suggest that future work should focus on quality control for slice selection to reduce these uncertainties. Additionally, while this study focused on 2D to save on radiation, the results highlight a strong argument for moving toward low-dose 3D (volumetric) analysis when tracking organ-specific health markers over time.

Summary updated

Abstract

Metabolic health is increasingly implicated as a risk factor across conditions from cardiology to neurology, and efficiency assessment of body composition is critical to quantitatively characterizing these relationships. 2D low dose single slice computed tomography (CT) provides a high resolution, quantitative tissue map, albeit with a limited field of view. Although numerous potential analyses have been proposed in quantifying image context, there has been no comprehensive study for low-dose single slice CT longitudinal variability with automated segmentation. We studied a total of 1816 slices from 1469 subjects of Baltimore Longitudinal Study on Aging (BLSA) abdominal dataset using supervised deep learning-based segmentation and unsupervised clustering method. 300 out of 1469 subjects that have two year gap in their first two scans were pick out to evaluate longitudinal variability with measurements including intraclass correlation coefficient (ICC) and coefficient of variation (CV) in terms of tissues/organs size and mean intensity. We showed that our segmentation methods are stable in longitudinal settings with Dice ranged from 0.821 to 0.962 for thirteen target abdominal tissues structures. We observed high variability in most organ with ICC&lt;0.5, low variability in the area of muscle, abdominal wall, fat and body mask with average ICC≥0.8. We found that the variability in organ is highly related to the cross-sectional position of the 2D slice. Our efforts pave quantitative exploration and quality control to reduce uncertainties in longitudinal analysis.
BibTeX
@inproceedings{yu2023longitudinal,
  author = {Xin Yu and Yucheng Tang and Qi Yang and Ho Hin Lee and Riqiang Gao and Shunxing Bao and Ann Zenobia Moore and Luigi Ferrucci and Bennett A. Landman},
  title = {{Longitudinal variability analysis on low-dose abdominal CT with deep learning-based segmentation}},
  booktitle = {Image Processing},
  year = {2023},
  doi = {10.1117/12.2653762},
}
Figure