You can deploy this model and start earning money today!

tan200224/
conditional-diffusion-medical

pytorch

diffusion

mit

Views

No views yet

Conditional Diffusion Model for Medical Image Generation

This repository contains a conditional diffusion model trained to generate 3D medical CT scan images based on segmentation masks.
The model uses a U-Net architecture with score-based diffusion for high-quality medical image synthesis.

Real or Fake Image?

Training Dataset

The model was trained on 3,346 CT scan examples with corresponding segmentation masks (80/20 train–validation split).

Sources:

Live Interactive Demo

Model Architecture

Base Model: U-Net with 5-level encoder–decoder
Input: 4-channel 256×256 CT scan images
Conditioning: Segmentation masks (4-channel 256×256)
Output: 4-channel 256×256 generated images
Sampling: Euler–Maruyama sampler (250 steps)
Training: Score matching loss with conditional generation

Model Details

Training Data: 3,346 CT scan examples
Lambda Parameter: 25.0 (diffusion coefficient)
Embedding Dimension: 256
Channels: [32, 64, 128, 256, 512]
Activation: SiLU (Swish)

Usage

This model can be used to add more diversity to your CT-scan dataset, especially when:

You have a limited dataset size (e.g., only a few hundred scans).
You want to balance underrepresented anatomical variations or rare conditions.
You need synthetic augmentation for training deep learning models in segmentation, detection, or classification.

Example Applications

Generate training samples from segmentation masks to reduce overfitting.
Create synthetic CT images with controlled variations to test robustness.
Improve representation of minority cases to reduce bias in medical AI.

Using the Hugging Face API

python

from transformers import AutoModelForImageGeneration
import torch

# Load the model
model = AutoModelForImageGeneration.from_pretrained("your-username/your-model-name")

# Generate images
conditioning_mask = torch.randn(1, 4, 256, 256)  # Your segmentation mask
generated_image = model.generate(conditioning_mask)


### Local Usage

```python
import torch
from model import UNet, marginal_prob_std, diffusion_coeff, Euler_Maruyama_sampler

# Load model
Lambda = 25.0
marginal_prob_std_fn = lambda t: marginal_prob_std(t, Lambda=Lambda, device='cuda')
score_model = UNet(marginal_prob_std=marginal_prob_std_fn)
score_model.load_state_dict(torch.load("ckpt_3D_v2.pth"))
score_model.eval()

# Generate sample
conditioning_mask = torch.randn(1, 4, 256, 256)
samples = Euler_Maruyama_sampler(
    score_model,
    marginal_prob_std_fn,
    lambda t: diffusion_coeff(t, Lambda=Lambda, device='cuda'),
    batch_size=1,
    x_shape=(4, 256, 256),
    num_steps=250,
    device='cuda',
    y=conditioning_mask
)

Training

The model was trained for 5000 epochs with:

Learning rate: 2e-4 (with decay)
Batch size: 1
Optimizer: Adam
Loss: Score matching loss

Citation

If you use this model in your research, please cite:

bibtex

@misc{conditional_diffusion_medical,
  title={Conditional Diffusion Model for Medical Image Generation},
  author={Archie Tan, Scott, Spurlock},
  year={2025},
  url={https://huggingface.co/tan200224}
}

Publications

Archie Tan, Scott Spurlock.
Learning to generate realistic medical images to improve pancreatic cancer segmentation.
Accepted for presentation at the 39th Annual Consortium for Computing Sciences in Colleges: Southeastern Conference (CCSC-SE 2025), Mercer University, Macon, GA, November 7–8, 2025.
Published in the Journal of the Consortium for Computing Sciences in Colleges (to appear).
Conference Site | Formatting Guidelines

License

This project is open-source under the MIT License.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Contact

For questions or issues, please open an issue on this repository.

tan200224/conditional-diffusion-medical