Semi-Supervised Standard-Dose PET Image Generation via Region-Adaptive Normalization and Structural Consistency...

Area of Science:

• Nuclear medicine
• Medical imaging
• Artificial intelligence in healthcare

Background:

• Positron Emission Tomography (PET) is crucial for clinical diagnosis but involves radiation exposure.
• Low-dose PET (LPET) reduces radiation but compromises image quality.
• Reconstructing high-quality Standard-dose PET (SPET) images from LPET data is essential for safe and effective imaging.

Purpose of the Study:

• To develop a novel approach for generating high-quality SPET images from LPET images.
• To enable safe reduction of tracer doses in PET imaging without sacrificing image quality.
• To leverage both paired and unpaired LPET and SPET datasets for improved reconstruction.

Main Methods:

• A semi-supervised deep learning framework was proposed for network training.
• Region-adaptive Normalization (RN) was introduced to handle intensity variations across different image regions.
• A structural consistency constraint was implemented to preserve anatomical details during image reconstruction.

Main Results:

• The proposed method successfully estimated high-quality SPET images from LPET data.
• Region-adaptive Normalization effectively mitigated the impact of intensity variations.
• The structural consistency constraint ensured the preservation of fine details in the reconstructed images.
• Experiments on real human chest-abdomen PET data demonstrated state-of-the-art performance.

Conclusions:

• The developed semi-supervised framework with RN and structural consistency offers an effective solution for low-dose PET reconstruction.
• This approach allows for reduced radiation exposure while maintaining diagnostic image quality.
• The method shows significant potential for improving the safety and efficacy of PET imaging in clinical practice.