Fast dynamic brain PET imaging using stochastic variational prediction for recurrent frame generation.

Area of Science:

• Neuroimaging
• Radiochemistry
• Artificial Intelligence

Background:

• Dynamic Positron Emission Tomography (PET) imaging provides crucial insights into tracer biodistribution in the brain.
• Acquiring dynamic PET data over extended periods can be time-consuming, potentially limiting patient comfort and throughput.
• Accurate prediction of later time frames from initial data could revolutionize dynamic PET acquisition protocols.

Purpose of the Study:

• To evaluate a recurrent frame generation algorithm for predicting late frames in dynamic brain PET imaging.
• To assess the algorithm's performance in reconstructing time-varying tracer biodistribution patterns.
• To determine the potential for reducing overall PET scan duration using predictive modeling.

Main Methods:

• A novel stochastic adversarial video prediction model was implemented.
• The model predicted the last 13 frames (25-90 minutes) from the initial 13 frames (0-25 minutes) of dynamic ¹⁸F-DOPA brain PET/CT studies.
• Quantitative analysis utilized Bland-Altman plots and region-wise Patlak graphical analysis on data from 46 subjects with ten-fold cross-validation.

Main Results:

• The algorithm successfully predicted the trend of time-varying tracer biodistribution.
• Lowest tracer uptake bias (-0.04) was observed in the putamen, with smallest variance in the cerebellum.
• Region-wise Patlak analysis showed average bias for $$ and distribution volume was approximately 4-5% in the caudate and putamen.

Conclusions:

• A novel deep learning approach was developed for accelerated dynamic brain PET imaging.
• The model can generate 65 minutes of time frames from an initial 25-minute acquisition.
• This method offers a significant reduction in scanning time, enhancing efficiency in neuroimaging studies.