Temporal information-guided dynamic dual-tracer PET signal separation network.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Dynamic dual-tracer positron emission tomography (PET) faces challenges in separating individual tracer information from mixed signals.
• Existing methods struggle with single-injection, single-scan dynamic dual-tracer PET image separation.

Purpose of the Study:

• To propose and evaluate a deep learning framework for separating dynamic dual-tracer PET images.
• To address the limitations of traditional methods in single-scan dual-tracer PET image analysis.

Main Methods:

• A novel deep learning framework utilizing a gated recurrent unit (GRU) network was developed.
• The network comprises encoder, separation (using bi-directional GRU), and decoder modules to process time activity curves (TACs).
• Performance was assessed using simulation data and realistic monkey PET data.

Main Results:

• The GRU-based network demonstrated superior performance with lower bias and mean squared error compared to stacked autoencoders and background subtraction in simulations.
• Realistic studies showed higher mean structural similarity and peak signal-to-noise ratio for the GRU network's results.
• The method effectively separates single-tracer information from dual-tracer PET data.

Conclusions:

• Temporal information-guided neural networks are feasible for single-injection, single-scan dynamic dual-tracer PET image separation.
• The GRU network leverages TAC temporal information, eliminating the need for arterial input functions (AIFs).
• This approach yields more robust and accurate separation results, significantly outperforming state-of-the-art methods.