The feasibility of K_i parametric imaging in clinical dynamic ¹⁸F-FDG total-body PET using a simulated-data-driven...

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Machine Learning

Background:

• Kinetic parameter (K_i) imaging offers high diagnostic accuracy but is limited by long scan times.
• Current machine learning methods for K_i imaging require substantial real-world data for training.

Purpose of the Study:

• To investigate a simulated-data-driven strategy for efficient clinical K_i parametric imaging with shortened scan durations.
• To overcome the data dependency of existing machine learning models for K_i image generation.

Main Methods:

• Generated simulated PET data with varying noise levels using the Patlak equation and K-Means clustering.
• Trained XGBoost models on simulated datasets to predict K_i values from short-duration (50-60 min) PET scans.
• Validated the approach using both simulated and real-world dynamic total-body PET data, comparing against the conventional Patlak method.

Main Results:

• Training with noise-inclusive simulated data significantly improved K_i value accuracy.
• The proposed method achieved superior performance over the conventional Patlak method on real-world data (Pearson's r=0.94 vs. 0.42, NMSE=0.11 vs. 5.33).
• Achieved higher peak signal-to-noise ratio (64.32 vs. 47.87) compared to the Patlak method.

Conclusions:

• A simulated-data-driven approach is feasible for generating reliable K_i images from clinical ¹⁸F-FDG dynamic total-body PET scans.
• This method effectively reduces the reliance on costly real-world data acquisition.
• Enables efficient K_i imaging with reduced scan times.