First deep learning framework for enhanced positron annihilation interaction-transmission imaging system precision

Positron annihilation interaction-transmission imaging (PAITI) is a novel low-dose, multiparameter anatomical imaging modality that utilizes positron annihilation photons to generate multiple 2D maps. PAITI produces secondary maps-interaction, deposited energy, attenuation, and electron density-from PAITI's basic maps via a stepwise analytical method, which currently yields an average relative error of 4.32%.

To improve precision for clinical applications (e.g., ion therapy treatment planning), we introduce DeepPAITI-the first deep learning approach in PAITI-to extract secondary maps with enhanced accuracy.

We developed a specialized deep learning architecture with a multibranch, multi-input, and multi-output framework. Extensive training and testing were performed on both numerical simulations and GATE Monte Carlo datasets. A sensitivity analysis was conducted, and DeepPAITI was compared against common models such as ResNet, UNet, and VGG-16.

DeepPAITI achieved mean relative errors (MREs) of 0.63  ±  0.08% for numerical simulations and 0.91  ±  0.20% for GATE-based tests, reflecting a 77% improvement over analytical methods. Under very low-dose conditions (< 1 µGy), electron density decomposition confidence increased from 63% to 99% for key materials (water, carbon, bone, and titanium). Additionally, DeepPAITI demonstrated at least an 83% improvement in MRE compared with conventional deep learning models.

DeepPAITI significantly surpasses traditional methods in extracting PAITI secondary maps, paving the way for more accurate clinical implementations in various field such as radiation and ion therapy treatment planning.