Authors: Dinglun He, Baoming Zhang, Xu Wang, Yao Hao, Deshan Yang, Ye Duan
Organizations: Department of Electrical Engineering and Computer Science, University of Missouri, Columbia, MO, USA · Department of Computer Science, The University of Texas at Dallas, Richardson, TX, USA · Department of Radiation Oncology, Washington University in St. Louis, MO, USA · Department of Radiation Oncology, Duke University, Durham, NC, USA · School of Computing, Clemson University, Clemson, SC, USA
Abdominal CT data are limited by high annotation costs and privacy constraints, which hinder the development of robust segmentation and diagnostic models. We present a Prior-Integrated Variation Modeling (PIVM) framework, a diffusion-based method for anatomically accurate CT image synthesis. Instead of generating full images from noise, PIVM predicts voxel-wise intensity variations relative to organ-specific intensity priors derived from segmentation labels. These priors and labels jointly guide the diffusion process, ensuring spatial alignment and realistic organ boundaries. Unlike latent-space diffusion models, our approach operates directly in image space while preserving the full Hounsfield Unit (HU) range, capturing fine anatomical textures without smoothing. Source code is available at https://github.com/BZNR3/PIVM.
Figures & tables
Figure 1: Variation Modeling via Diffusion for CT Image Synthesis. The variation is forward diffused to Gaussian noise. A reverse diffusion model, guided by the segmentation label and average intensity map, predicts the variation signal, which is added back to reconstruct the final CT image.
Figure 2: Construction of Average Intensity Map and Variation. An average intensity map is computed from the segmentation label by replacing each organ region with its dataset-level mean value. The pixel-wise difference between the original CT image and the average map defines the variation used in subsequent diffusion modeling.
Figure 3: Examples of PIVM’s generation process. Columns show the ground-truth CT image, the corresponding average organ intensity priors, the predicted intensity variations, and the reconstructed CT obtained by combining the prior with the variation.
Table 1: Comparison of SSIM and FID among CDDPM, MAISI, and PIVM.
Figure 5: Multi-view visualization of a generated 3D abdominal CT volume, illustrating anatomical fidelity and inter-slice continuity. Left: coronal view with overlaid multi-organ contours. Right: sagittal view of the same volume.
Metric
CDDPM
MAISI
PIVM (ours)
Hausdorff Distance (↓)
52.95
61.21
43.16
Average Surface Dist. (↓)
13.23
17.36
6.60
mIoU (↑)
0.30
0.27
0.48
Dice Coefficient (↑)
0.38
0.34
0.53
Table 2: Segmentation Performance Comparison for Models Trained on Datasets Generated by CDDPM, MAISI, and PIVM.
Figure 6: Image generation results under different conditions. From left to right: ground truths (GT), images generated from 34 labels , and images generated from 13 labels .
Metric
Original Dataset
Generated Dataset
Combined Dataset
Hausdorff Distance (↓)
23.54
43.16
22.24
mIoU (↑)
0.78
0.48
0.81
Dice Coefficient (↑)
0.87
0.53
0.88
Table 3: Segmentation Performance for Models Trained on Original Dataset, Generated Dataset, and Combined Dataset.
Center for Virtual Imaging Trials, Dept. of Radiology, Duke University School of Medicine · Dept. of Physics, Trinity School of Arts and Sciences, Duke University · Dept. of Electrical and Computer Engineering, Pratt School of Engineering, Duke University
Mallinckrodt Institute of Radiology, Washington University School of Medicine, St. Louis, USA · Imaging Science Program, McKelvey School of Engineering, Washington University in St Louis, St. Louis, USA · Department of Surgery, Washington University School of Medicine, St. Louis, USA +1