PIVM: Diffusion-Based Prior-Integrated Variation Modeling for Anatomically Precise Abdominal CT Synthesis

PIVM: Diffusion-Based Prior-Integrated Variation Modeling for Anatomically Precise Abdominal CT Synthesis Dinglun He*, Baoming Zhang*, Xu Wang*, Yao Hao, Deshan Yang, Ye Duan (* Equal contribution) IEEE International Symposium on Biomedical Imaging (ISBI), 2026 [Paper] | [Data]

---

Overview

PIVM is a diffusion-based generative model for volumetric medical image synthesis. It learns to reconstruct CT image slices conditioned on organ segmentation masks and intensity-based label maps, using a residual learning strategy within a denoising diffusion probabilistic framework.

Key features:

• Residual diffusion learning: the model predicts the residual between a CT image and its organ-intensity label, rather than raw pixel values
• Sequential volume generation: slices are generated one by one, each conditioned on the previous generated slice
• Multi-modal conditioning: combines organ masks, intensity labels, and previous-slice context

---

Requirements

conda env create -f environment.yml
conda activate pivm

Tested with Python 3.9, PyTorch 1.12.1, CUDA 11.3.

---

Data

Download: https://zenodo.org/records/10047292

After downloading, the raw dataset follows the TotalSegmentator format:

Totalsegmentator_dataset/
└── s0001/
    ├── ct.nii.gz
    └── segmentations/
        ├── liver.nii.gz
        ├── spleen.nii.gz
        ├── pancreas.nii.gz
        └── ...

---

Data Preprocessing

All preprocessing scripts are located in preprocessing/. Run them in order.

Step 1 — Merge organ segmentations into a single mask

preprocessing/ts_process.py combines individual per-organ NIfTI files into a single multi-label mask.nii per subject.

python preprocessing/ts_process.py

Step 2 — Convert NIfTI volumes to 2D PNG slices

preprocessing/vol2imglabel.py converts each subject's CT volume and merged mask into 2D PNG slices.

python preprocessing/vol2imglabel.py

Step 3 — Generate intensity-based conditioning labels

preprocessing/process_ct_intensity.py replaces each organ region in the mask with its global mean CT intensity across the full dataset, producing the conditioning label maps used during training.

python preprocessing/process_ct_intensity.py \
  --ct_dir    ./image \
  --label_dir ./label \
  --output_dir ./data/train/label

---

Training

Set training = True, configure the paths inside the script, then run:

python train_test_ddpm.py

To resume from a checkpoint, set load_model = True and point checkpoint_path to your saved checkpoint.

Training outputs are saved to:

results/
├── checkpoints/       # Model checkpoints (saved when loss improves)
└── *.png              # Intermediate sample visualizations

---

Testing

Set training = False, configure the paths inside the script, then run:

python train_test_ddpm.py

This runs sequential volume reconstruction — slices are generated one by one, each conditioned on the previous generated slice. Results are saved to results/.

---

Citation

@article{pivm,
  title={PIVM: Diffusion-Based Prior-Integrated Variation Modeling for Anatomically Precise Abdominal CT Synthesis},
  author={He, Dinglun and Zhang, Baoming and Wang, Xu and Hao, Yao and Yang, Deshan and Duan, Ye},
  booktitle={International Symposium on Biomedical Imaging (ISBI)},
  year={2026}
}