Deep conditional generative models for longitudinal single-slice abdominal computed tomography harmonization

Xin Yu, Qi Yang, Yucheng Tang, Riqiang Gao, Shunxing Bao, Leon Y. Cai, Ho Hin Lee, Yuankai Huo, Ann Zenobia Moore, Luigi Ferrucci, Bennett A. LandmanarXiv 2023 · DOI · OpenReview

Introduction

Longitudinal medical imaging studies are essential for tracking biological changes over time, particularly in aging and metabolic health. However, a significant challenge in these studies is data inconsistency. This research introduces C-SliceGen, a deep conditional generative model designed to "harmonize" 2D abdominal computed tomography (CT) scans. By standardizing arbitrary 2D slices to a consistent vertebral level, the model enables more accurate tracking of body composition biomarkers like visceral fat and muscle area across multiple years.

The Problem: Positional Variance in 2D CT

In large-scale longitudinal studies, researchers often utilize single-slice 2D CT scans rather than full 3D volumes to minimize radiation exposure. However, it is anatomically difficult for technicians to capture the exact same vertebral level year after year. Even a 2cm difference in slice position can significantly alter the visible amount of visceral fat or muscle, creating "positional variance" that masks true biological trends. Traditional image registration techniques are often ineffective here because 2D slices lack the necessary "out-of-plane" information to adjust the anatomical level.

Longitudinal 2D CT scans across three visits showing positional variance and visceral fat segmentation.

Methodology: C-SliceGen

The proposed solution, C-SliceGen, is a hybrid model based on a Variational Autoencoder (VAE) and a Generative Adversarial Network (GAN). The architecture is designed to learn a latent representation that separates subject-specific characteristics from anatomical positioning.

Model Architecture

The model employs a dual-encoder system:

  1. Encoder 1: Captures subject-specific details (e.g., body habitus) from an arbitrary input slice.
  2. Encoder 2: Learns the target anatomical structure from a reference slice during training.
  3. Decoder: Synthesizes the harmonized image by combining the subject-specific latent code with the target anatomical code.
The C-SliceGen architecture, a conditional VAEGAN designed for slice harmonization.

To ensure high-quality results, the model utilizes a combination of loss functions: reconstruction loss (L1-norm) for accuracy, KL-divergence for latent space regularization, and a GAN loss (WGAN-GP) to ensure the generated images maintain the sharpness and texture of real CT scans.

Experimental Setup and Datasets

The model was trained and validated using a diverse set of CT data to ensure robustness across different imaging conditions:

  • Portal Venous & Non-contrast Datasets: Over 2,600 3D CT volumes were used to train the model and address domain shifts between contrast-enhanced and non-contrast scans.
  • BTCV Challenge: 50 subjects were used for external validation of the model's generalization capabilities.
  • Baltimore Longitudinal Study of Aging (BLSA): The primary application target, consisting of 1,033 subjects with longitudinal 2D slices collected over a 15-year period.

Key Insights and Results

C-SliceGen demonstrated a significant ability to reduce noise in longitudinal data while preserving critical biological information:

  • Variance Reduction: In the BLSA dataset, the model reduced the coefficient of variation (CV) in subjects with misaligned scans by 36.3% to 76.9%, effectively smoothing out the noise caused by inconsistent positioning.
  • Biomarker Preservation: Beyond visual fidelity, the model successfully harmonized the visceral fat area (VFA), a key metric for metabolic health, ensuring that changes observed over time were biological rather than procedural.
  • Optimal Generation Range: Ablation studies revealed that the model is most effective when the input slice is between 15mm and 60mm from the target level. Beyond this range, structural changes become too drastic for accurate prediction.

Conclusion and Future Directions

C-SliceGen provides a powerful tool for standardizing legacy longitudinal datasets where 3D volumes are unavailable. While the model excels at harmonizing fat and muscle, future work will focus on improving the generation of highly variable organs like the stomach and colon, which change shape based on posture and digestion. Additionally, implementing a "gatekeeper" mechanism to determine when harmonization is strictly necessary could further refine the pipeline's accuracy.

Summary updated

Abstract

Two-dimensional single-slice abdominal computed tomography (CT) provides a detailed tissue map with high resolution allowing quantitative characterization of relationships between health conditions and aging. However, longitudinal analysis of body composition changes using these scans is difficult due to positional variation between slices acquired in different years, which leading to different organs/tissues captured. To address this issue, we propose C-SliceGen, which takes an arbitrary axial slice in the abdominal region as a condition and generates a pre-defined vertebral level slice by estimating structural changes in the latent space. Our experiments on 2608 volumetric CT data from two in-house datasets and 50 subjects from the 2015 Multi-Atlas Abdomen Labeling Challenge dataset (BTCV) Challenge demonstrate that our model can generate high-quality images that are realistic and similar. We further evaluate our method's capability to harmonize longitudinal positional variation on 1033 subjects from the Baltimore Longitudinal Study of Aging (BLSA) dataset, which contains longitudinal single abdominal slices, and confirmed that our method can harmonize the slice positional variance in terms of visceral fat area. This approach provides a promising direction for mapping slices from different vertebral levels to a target slice and reducing positional variance for single-slice longitudinal analysis. The source code is available at: https://github.com/MASILab/C-SliceGen.
BibTeX
@misc{yu2023deep,
  author = {Xin Yu and Qi Yang and Yucheng Tang and Riqiang Gao and Shunxing Bao and Leon Y. Cai and Ho Hin Lee and Yuankai Huo and Ann Zenobia Moore and Luigi Ferrucci and Bennett A. Landman},
  title = {{Deep conditional generative models for longitudinal single-slice abdominal computed tomography harmonization}},
  year = {2023},
  eprint = {2309.09392},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2309.09392},
}
Figure