Three-dimensional convolutional neural networks for simultaneous dual-tracer PET imaging.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence in Healthcare

Background:

• Dual-tracer positron emission tomography (PET) enables simultaneous measurement of two tracers in a single scan, enhancing diagnostic accuracy and research capabilities.
• Current dual-tracer PET reconstruction methods often rely on pre-reconstructed mixed images, limiting reconstruction precision.
• Developing advanced reconstruction algorithms is crucial for improving the accuracy and utility of dual-tracer PET imaging.

Purpose of the Study:

• To present a hybrid loss-guided deep learning framework for accurate dual-tracer PET imaging directly from sinogram data.
• To unify image reconstruction and tracer separation within a single framework.
• To evaluate the performance and robustness of the proposed method compared to existing deep learning techniques.

Main Methods:

• A hybrid loss-guided deep learning framework utilizing a 3D convolutional neural network (CNN) was developed for dual-tracer PET reconstruction from sinogram data.
• The framework integrates reconstruction of mixed images and separation of individual tracers using a unified approach.
• Monte Carlo simulations with data augmentation were employed for training and testing, with performance assessed using bias and variance analysis across spatial regions and temporal frames.

Main Results:

• The 3D CNN framework demonstrated feasibility for accurate dual-tracer PET reconstruction, effectively learning spatial and temporal features.
• The proposed method showed superior performance compared to a deep belief network (DBN) in dual-tracer image separation.
• The framework exhibited robustness to noise, successfully recovering tracer distributions with high accuracy even at lower total counts.

Conclusions:

• The hybrid loss-guided deep learning framework offers a significant advancement in dual-tracer PET imaging reconstruction and separation.
• The 3D CNN approach provides enhanced precision and robustness, outperforming traditional deep learning methods like DBN.
• This method holds promise for improving clinical diagnosis and advancing scientific research using dual-tracer PET.