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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy Physics

Background:

• Positron Annihilation Interaction-Transmission Imaging (PAITI) is a low-dose imaging technique generating multiple 2D maps.
• Current analytical methods for extracting secondary maps (e.g., electron density) have an average relative error of 4.32%.

Purpose of the Study:

• To introduce DeepPAITI, the first deep learning approach for PAITI, to enhance the accuracy of secondary map extraction.
• Improve precision for clinical applications such as ion therapy treatment planning.

Main Methods:

• Developed a specialized deep learning architecture with a multibranch, multi-input, and multi-output framework.
• Trained and tested the model using numerical simulations and GATE Monte Carlo datasets.
• Compared DeepPAITI against conventional models like ResNet, UNet, and VGG-16.

Main Results:

• DeepPAITI achieved mean relative errors (MREs) of 0.63% (simulations) and 0.91% (GATE tests), a 77% improvement over analytical methods.
• Electron density decomposition confidence increased from 63% to 99% under low-dose conditions (< 1 µGy).
• Demonstrated at least an 83% improvement in MRE compared to conventional deep learning models.

Conclusions:

• DeepPAITI significantly surpasses traditional and conventional deep learning methods for PAITI secondary map extraction.
• Paves the way for more accurate clinical implementations in radiation and ion therapy treatment planning.