AI Researcher working on Computer Vision and Medical Imaging.

DACD: Dose-Aware Cold Diffusion with Physics Consistency for Generalizable Low-Dose CT Reconstruction

Lower-dose CT scans are safer for patients but produce noisier images; DACD reconstructs clean images from noisy low-dose scans more accurately and 3.3× faster than prior methods.

+1.20 dB PSNR over SOTA, 3.3× faster inference.

Problem

Standard diffusion-based CT reconstruction methods are accurate but slow, and often fail to generalize across the wide range of dose levels used in clinical practice.

Method

A cold diffusion framework built on Poisson thinning — modeling the physical noise process of low-dose acquisition directly — combined with an explicit physics-consistency constraint tying the reconstruction back to the measured sinogram.

Contribution

Reformulates the degradation process in cold diffusion around the actual physics of dose-reduced CT acquisition, rather than treating noise as a generic corruption to invert.

• PyTorch
• Cold Diffusion
• CT/Sinogram Processing

Intl. Joint Conf. on Neural Networks (IJCNN 2026)

GenDiff: A Dose and Anatomy Aware Diffusion Model with Structural Prior Refinement for Low-Dose CT Reconstruction and Generalization

Extends the DACD line of work so a single trained model reconstructs CT images well across dose levels and body regions it was never explicitly trained on.

Problem

A reconstruction model trained at one dose level or anatomical region typically degrades sharply when deployed on a different dose level or body region — a real barrier to clinical deployment.

Method

Explicit dose and anatomy conditioning inside the diffusion process, paired with a structural prior refinement stage that corrects fine anatomical detail after the main reconstruction step.

Contribution

Demonstrates zero-shot generalization across dose levels and anatomical domains without per-setting retraining.

• PyTorch
• Diffusion Models
• Domain Generalization

PeerJ Computer Science

LightGAN-LD: A Lightweight Generative Adversarial Network for Efficient Low-Dose CT Reconstruction with Sinogram Encoding and Edge-Aware Learning

A smaller, faster generative model for low-dose CT reconstruction, built for real-time deployment rather than offline processing.

Problem

Diffusion-based reconstruction is accurate but computationally heavy; clinical real-time use cases need a much lighter model without giving up reconstruction quality or edge fidelity.

Method

A lightweight GAN architecture operating directly on sinogram-encoded input with an edge-aware learning objective to preserve diagnostically relevant structural detail.

Contribution

Trades diffusion-model accuracy for GAN-level inference speed while retaining edge fidelity, in an international collaboration with Dr. Mohammad Azam Khan at KAIST. A closely related sinogram-encoded variant targeting real-time measurement-domain noise suppression is under review as a separate paper (R4).

• PyTorch
• GANs
• Sinogram Processing

PeerJ Computer Science