Introduction
Quantitative analysis of body composition—specifically the measurement of fat, muscle, and bone—is a critical component of aging and metabolic research. While 3D CT volumes provide comprehensive data, 2D single-slice CT imaging is frequently utilized in clinical and research settings to minimize radiation exposure and reduce computational overhead.
However, longitudinal studies face a significant challenge: positional variance. It is technically difficult to capture a 2D slice at the exact same anatomical landmark (e.g., the midpoint of the L3 vertebra) across multiple visits years apart. Because abdominal anatomy changes rapidly along the vertical axis, even minor deviations in slice location can introduce "artificial" noise into body composition metrics, making it difficult to distinguish true biological change from measurement error. To address this, researchers developed C-SliceGen, a deep conditional generative model designed to synthesize standardized 2D abdominal CT slices from arbitrary starting positions.
Methodology
C-SliceGen utilizes a VAEGAN architecture—a hybrid of a Variational Autoencoder (VAE) and a Generative Adversarial Network (GAN). The model is designed to map varied longitudinal scans to a consistent vertebral level by extracting subject-specific features and combining them with structural information about a target level in a latent space.
Architecture and Training
The model employs two encoders and one decoder. During training, one encoder processes a "conditional" input slice while the second processes a "target" slice. The model learns to disentangle body habitus from anatomical level. During inference, the model requires only the arbitrary conditional slice; it samples target-level information from a learned prior to synthesize the standardized version. The optimization process balances reconstruction loss ($L_1$), KL-divergence for latent space regularization, and an adversarial GAN loss ($\beta L_{GAN}$) to ensure the generated images maintain high anatomical realism.
Experimental Setup
The study validated the model using three distinct datasets:
- In-house Dataset: 1,170 3D Portal Venous CT volumes for primary training.
- BTCV MICCAI Challenge 2015: 50 subjects used for external validation of image quality.
- Baltimore Longitudinal Study of Aging (BLSA): 20 subjects with longitudinal data spanning up to 15 years, used to test the model's ability to harmonize noisy longitudinal measurements.
Results and Discussion
The inclusion of GAN loss significantly improved the perceptual quality of the synthesized slices. On the BTCV dataset, the model achieved an LPIPS score of 0.208 (where lower is better), a marked improvement over the 0.362 score achieved without GAN regularization.
Longitudinal Harmonization
The most significant impact of C-SliceGen was observed in longitudinal stability. By "moving" all slices from different visits to a virtual standardized level, the researchers reduced the volatility of muscle and fat area measurements. This harmonization allows for more plausible tracking of biological trends over time, smoothing out the fluctuations caused by inconsistent slice positioning during acquisition.
Conclusion and Future Work
C-SliceGen provides a "plug-and-play" solution for researchers to stabilize 2D CT data without requiring manual ground-truth labels for every slice. While the model successfully shifts anatomy—such as generating kidneys where they were previously absent—it still faces challenges with highly variable soft tissues like the stomach and colon. Future work will focus on improving robustness across different CT acquisition protocols and enhancing the synthesis of heterogeneous soft tissues to further refine longitudinal body composition analysis.
Summary updated