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

2026-07-11 · Source: Computer Vision and Pattern Recognition · Field: Science & Research — Health & Medical Research, Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Expert, quick

Summary

GenDiff is a novel diffusion-based framework designed for low-dose computed tomography (LDCT) reconstruction, addressing the limitations of existing methods that struggle with varying radiation dose levels and anatomical regions. This framework jointly models continuous radiation dose and anatomical information within a unified network. Key components include a Dose-Anatomy Encoder for acquisition-aware embeddings, a dose- and anatomy-conditioned cold diffusion backbone for iterative refinement, a physics-consistency update, and a Structural Prior Refinement Module (SPRM) to preserve anatomical structures. Extensive experiments on multi-anatomy clinical datasets, including unseen ultra-low-dose conditions and out-of-distribution phantom/animal datasets, demonstrate GenDiff's superior performance over state-of-the-art convolutional neural network and diffusion-based methods. It achieves robust reconstruction quality across diverse clinical scenarios.

Key takeaway

For Machine Learning Engineers developing low-dose CT reconstruction models, GenDiff's approach of jointly modeling continuous radiation dose and anatomical information offers a robust solution. You should consider integrating dose-anatomy encoders and structural prior refinement to enhance generalization across diverse clinical datasets and ultra-low-dose conditions, improving diagnostic image quality. This method promises superior performance and robustness in practical clinical settings.

Key insights

GenDiff integrates dose and anatomy awareness into a diffusion model for robust low-dose CT reconstruction.

Principles

• Jointly model dose and anatomy for generalization.
• Refine structural priors to suppress artifacts.
• Enforce physics-consistency for data fidelity.

Method

GenDiff employs a Dose-Anatomy Encoder, a conditioned cold diffusion backbone, physics-consistency updates, and a Structural Prior Refinement Module (SPRM).

In practice

• Reconstruct multi-anatomy clinical CT scans.
• Process ultra-low-dose imaging conditions.
• Generalize to out-of-distribution datasets.

Topics

• Low-Dose CT Reconstruction
• Diffusion Models
• Medical Imaging
• Generalization
• Radiation Dose
• Anatomy-Aware AI

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Supervise-assisted Multi-modality Fusion Diffusion Model for PET Restoration — MFdiff restores high-quality PET from low-dose inputs using multi-modality fusion and a two-stage supervised diffusion model.
• Scaling Generative Foundation Models for Chest Radiography with Rectified Flow Transformers — A 1.3B-parameter generative foundation model for chest radiographs achieves expert-indistinguishable synthesis and controllable editing.
• Treatment-Conditioned Diffusion for Forecasting Neurodegenerative Disease Progression — A treatment-conditioned diffusion framework accurately forecasts neurodegenerative disease progression by integrating imaging and pharmacological data, preserving anatomical detail.
• NAMD: Virtual Follow-up Computed Tomography Synthesis via Nodule-Aligned Multimodal Diffusion Models for Early Lung Cancer Diagnosis — NAMD synthesizes future lung nodule CTs from baseline and EHR using aligned latent diffusion for early cancer diagnosis.
• LUCID: Learned Undersampling-Adaptive Consistency-Guided Inference with Deterministic Flow Matching for Sparse-View CT Reconstruction — LUCID employs Flow Matching and sparsity-adaptive consistency guidance for robust sparse-view CT reconstruction, reducing artifacts and hallucinations.
• CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training — Latent diffusion with RL post-training significantly enhances controllable 3D medical image synthesis fidelity and faithfulness.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.