LPGNet: Learnable Prompt-Guided Network for Low-dose Contrast Computed Tomography Angiography Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Iodinated contrast agents (ICAs) are crucial for computed tomography angiography (CTA) but pose risks like contrast-induced nephropathy.
• Low-dose contrast agent CTA (LDCTA) offers a safer alternative but faces challenges in maintaining vessel intensity and boundary definition.

Purpose of the Study:

• To develop an advanced framework, LPGNet, for unpaired vascular reconstruction in LDCTA.
• To improve the quality and diagnostic utility of low-dose contrast vascular imaging.

Main Methods:

• LPGNet integrates Contrastive Language-Image Pretraining (CLIP) with anatomical intensity consistency constraints.
• Learnable prompts guide CLIP for contrast-specific semantics, while intensity loss ensures optimal aortic enhancement.
• An iterative optimization strategy refines the network and prompts via semantic alignment.

Main Results:

• LPGNet effectively differentiates low-contrast vessels and identifies fine-grained structural details.
• The framework enhances vascular contrast and preserves anatomical details in LDCTA.
• Experiments show LPGNet outperforms existing paired and unpaired methods on simulated and clinical data.

Conclusions:

• LPGNet provides a robust solution for contrast-limited vascular imaging.
• The method significantly improves diagnostic utility in LDCTA.
• LPGNet demonstrates potential for safer and more effective vascular imaging.