Visual language model-assisted CT denoising via text-guided diffusion and fidelity maintenance.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Reducing radiation dose in computed tomography (CT) and photon-counting CT (PCCT) is vital for patient safety.
• Low-dose CT imaging introduces noise, degrading image quality and potentially impacting diagnostic accuracy.
• Current denoising methods often require paired data or rely on specific noise assumptions, limiting clinical applicability.

Purpose of the Study:

• To introduce a novel Visual-Language Model-assisted CT Denoising (VLD) framework for effective CT image noise reduction.
• To leverage semantic understanding from visual-language models for improved CT image restoration.
• To preserve diagnostic fidelity and structural integrity in low-dose CT images.

Main Methods:

• Developed a Visual-Language Model-assisted CT Denoising (VLD) framework utilizing semantic guidance.
• Employed a diffusion model guided by semantic understanding derived from multimodal visual-language models.
• Implemented a tri-domain consistency framework for progressive refinement of image details and structural preservation.

Main Results:

• The VLD method demonstrated high-quality reconstruction on simulated CT and real PCCT data.
• Achieved average peak signal-to-noise ratio improvements of 0.95 dB and 1.21 dB under specific conditions.
• Outperformed existing methods like WGAN and FBPConvNet, which require paired data, in simulation experiments.
• Showcased robust generalization capabilities to new imaging scenarios.

Conclusions:

• The VLD framework effectively reduces noise in CT images while maintaining diagnostic quality.
• Leveraging visual-language models offers a promising direction for advanced medical image denoising.
• The proposed method provides a robust and generalizable solution for low-dose CT denoising, overcoming limitations of previous approaches.