A generative adversarial network to improve integrated mode proton imaging resolution using paired proton-carbon...

Area of Science:

• Medical Imaging
• Particle Physics
• Artificial Intelligence

Background:

• Proton radiography (pRad) offers clinical accessibility but suffers from limited spatial resolution due to multiple Coulomb scattering (MCS).
• Heavier ions like carbon ions exhibit reduced MCS, leading to inherently higher resolution radiographs (cRads).
• Image resolution enhancement for pRads can be achieved by translating proton data to equivalent carbon ion data via image translation networks.

Purpose of the Study:

• To develop and evaluate a deep learning framework for increasing the spatial resolution of integrated mode proton radiographs.
• To leverage paired proton-carbon imaging data for high-fidelity image translation.

Main Methods:

• A conditional generative adversarial network, termed Proton2Carbon, was developed for translating proton pencil beam images to synthetic carbon ion beam images.
• The model was trained on a large dataset of 547,224 paired proton-carbon images acquired using a scintillation detector.
• Evaluation involved assessing spatial resolution using custom 3D-printed line pair modules on internal and external datasets.

Main Results:

• The Proton2Carbon model successfully improved pRad spatial resolution from 1.7 to 2.7 lp/cm (internal) and 2.3 lp/cm (external), demonstrating generalizability.
• Water equivalent thickness accuracy was maintained, comparable to both pRads and cRads.
• Phantom studies revealed enhanced structural clarity in translated images, despite a slight increase in observed noise.

Conclusions:

• Deep learning effectively enhances proton radiography image quality by utilizing paired proton-carbon data.
• The Proton2Carbon framework shows potential for integration into clinical imaging workflows, improving proton radiography applications.
• The training dataset is publicly released to encourage further research in this domain.