Deep learning-based dynamic PET parametric K_i image generation from lung static PET

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• Positron Emission Tomography/Computed Tomography (PET/CT) is crucial for lung cancer diagnosis.
• Static PET imaging has limitations in quantification accuracy.
• Dynamic PET parametric K_i imaging offers improved quantification and cancer detection specificity but requires long acquisition times.

Purpose of the Study:

• To develop a deep learning-based method for synthesizing dynamic parametric K_i images from conventional static PET scans.
• To overcome the clinical limitations of long acquisition times for dynamic PET imaging.

Main Methods:

• Utilized data from 203 participants.
• Developed an improved cycle generative adversarial network with a squeeze-and-excitation attention block.
• Trained the network to map static PET images to K_i parametric images.
• Evaluated synthesized images using qualitative, quantitative, and statistical analyses.

Main Results:

• The proposed deep learning network synthesized K_i images with superior performance compared to other networks.
• Synthesized K_i images showed high correlation (Pearson correlation coefficient, 0.93) and consistency with standard dynamic PET.
• Excellent quantitative evaluation results were achieved for the synthesized images.

Conclusions:

• The deep learning method effectively synthesizes dynamic parametric images from static lung PET.
• This approach provides a valuable, quantitative diagnostic reference for clinicians.
• Enables improved lung cancer detection and characterization without extended scan times.