Combining deep learning with a kinetic model to predict dynamic PET images and generate parametric images

Area of Science:

• Medical Imaging
• Radiochemistry
• Artificial Intelligence in Medicine

Background:

• Dynamic PET imaging enables metabolic parameter (K_i) calculation for clinical diagnosis.
• Current 1-hour scan times limit dynamic PET utility due to trade-offs between scan duration and K_i image signal-to-noise ratios (SNRs).

Purpose of the Study:

• To develop a method for generating high-quality dynamic PET images comparable to 1-hour scans within a 30-minute timeframe.
• To improve SNRs of 30-minute dynamic PET scans and reduce overall scanning duration.

Main Methods:

• Utilized a U-Net architecture as a feature extractor for image structure.
• Employed a parameter generator to create a time activity curve (TAC) for a two-tissue, three-compartment model.
• Generated K_i parameter images from the synthesized dynamic PET data.

Main Results:

• Network-generated K_i maps showed improved structural similarity index measure (2.27%) and peak SNR (7.04%) compared to standard 30-min scans.
• Root mean square error (RMSE) was reduced by 16.3% for the deep learning-derived K_i maps.
• The method successfully generated comparable K_i parameter images in half the standard scanning time.

Conclusions:

• The proposed deep learning method is feasible for achieving satisfactory PET quantification accuracy.
• This approach offers potential for reduced scanning times in dynamic PET imaging.
• Further clinical validation is required for routine application.