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

Background: Integrated mode proton imaging is a clinically accessible method for proton radiographs (pRads), but its spatial resolution is limited by multiple Coulomb scattering (MCS). As the amplitude of MCS decreases with increasing particle charge, heavier ions such as carbon ions produce radiographs with better resolution (cRads). Improving image resolution of pRads may thus be achieved by transferring individual proton pencil beam images to the equivalent carbon ion data using a trained image translation network. The approach can be interpreted as applying a data-driven deconvolution operation with a spatially variant point spread function.

Purpose: Propose a deep learning framework based on paired proton-carbon data to increase the resolution of integrated mode pRads.

Methods: A conditional generative adversarial network, Proton2Carbon, was developed to translate proton pencil beam images into synthetic carbon ion beam images. The model was trained on 547 224 paired proton-carbon images acquired with a scintillation detector at the Marburg Ion Therapy Centre. Image reconstruction was performed using a 2D lateral method, and the model was evaluated on internal and external datasets for spatial resolution, using custom 3D-printed line pair modules.

Results: The Proton2Carbon model improved the spatial resolution of pRads from 1.7 to 2.7 lp/cm on internal data and to 2.3 lp/cm on external data, demonstrating generalizability. Water equivalent thickness accuracy remained consistent with pRads and cRads. Evaluation on an anthropomorphic head phantom showed enhanced structural clarity, though some increased noise was observed.

Conclusions: This study demonstrates that deep learning can enhance pRad image quality by leveraging paired proton-carbon data. Proton2Carbon can be integrated into existing imaging workflows to improve clinical and research applications of proton radiography. To facilitate further research, the full dataset used to train Proton2Carbon is publicly released and available at https://zenodo.org/records/14945165.