FoundDiff: Foundational Diffusion Model for Generalizable Low-Dose CT Denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) denoising is vital for reducing radiation exposure while maintaining diagnostic image quality.
• Current deep learning (DL) methods lack generalizability due to training on specific dose levels and anatomical regions, limiting clinical application.
• Diverse noise characteristics and anatomical variations in varied scanning conditions pose challenges for existing denoising techniques.

Purpose of the Study:

• To propose FoundDiff, a unified and generalizable foundational diffusion model for LDCT denoising.
• To address the limitations of existing DL methods in handling diverse noise and anatomical heterogeneity.
• To achieve robust and adaptive denoising across various dose levels and anatomical regions.

Main Methods:

• A two-stage strategy involving dose-anatomy perception and adaptive denoising.
• Development of a dose- and anatomy-aware contrastive language-image pre-training model (DA-CLIP) for robust perception of dose variations and anatomical regions.
• Design of a dose- and anatomy-aware diffusion model (DA-Diff) integrating learned embeddings via a novel Mamba-based conditional block (DACB) for adaptive denoising.

Main Results:

• Extensive experiments on simulated and public datasets (Mayo-2016, CQ500, piglet) demonstrate superior denoising performance.
• FoundDiff shows strong generalization capabilities to unseen dose levels and anatomical regions.
• The proposed method achieves unified and generalizable LDCT denoising.

Conclusions:

• FoundDiff offers a robust and generalizable solution for LDCT denoising, overcoming limitations of previous methods.
• The model's ability to adapt to various dose levels and anatomical regions enhances its clinical utility.
• This work advances the field of medical image denoising with a foundational diffusion model approach.