Physics-Informed Diffusion Modelling for Medical Imaging

An ongoing project at the SDAIA-KFUPM Joint Research Center for AI, developing physics-guided diffusion architectures that incorporate imaging operators (e.g., point spread function, attenuation, noise models) to improve reconstruction fidelity and robustness in medical imaging.

The work focuses on making these generative models practical for deployment:

Physics-guided diffusion architectures that explicitly model the imaging forward operator.
Computationally efficient diffusion via few-step sampling, lightweight backbones, knowledge distillation, and quantization for real-time and edge deployment.
Integrating generative outputs with downstream tasks (segmentation, classification, diagnostics) for task-aware performance improvements.

This work builds on related efficient medical image segmentation research, including (Munir et al., 2026).

References

2026

JBHI

DAUNet: A lightweight UNet variant with deformable convolutions and parameter-free attention for medical image segmentation

Adnan Munir, Muhammad Shahid Jabbar, and Shujaat Khan

IEEE Journal of Biomedical and Health Informatics, 2026

Bib PDF

@article{munir2026daunet,
  title = {DAUNet: A lightweight UNet variant with deformable convolutions and parameter-free attention for medical image segmentation},
  author = {Munir, Adnan and Jabbar, Muhammad Shahid and Khan, Shujaat},
  journal = {IEEE Journal of Biomedical and Health Informatics},
  year = {2026},
  publisher = {IEEE},
}