Author: Mostafa Cham, Matthias K Gobbert, Zhuoran Jiang, Sina Mossahebi, Ruth Obe, Stephen W. Peterson, Jerimy C. Polf, Lei Ren, Ehsan Shakeri, Vijay Raj Sharma ๐จโ๐ฌ
Affiliation: University of Maryland School of Medicine, UMBC, University of Maryland Baltimore County, University of Maryland, Baltimore County, Stanford University, University of Maryland, School of Medine, Department of Physics, University of Cape Town, M3D, Inc, Department of Infomation Systems, UMBC ๐
Purpose: Compton camera (CC)-based prompt gamma imaging (PGI) offers real-time proton range verification. However, its limited-angle measurements cause severe distortions in PGI, affecting its clinical utility. Previously, we have demonstrated the efficacy of deep learning (DL) to address these issues using a small dataset from water phantoms and patient anatomy. This study aims to optimize the DL modelโs architecture and further validate its efficacy using a large dataset simulated from patient data.
Methods: An attention-driven two-tier DL framework was developed for PGI reconstruction. The first tier includes a localization model defining arbitrary-shaped region-of-interest (ROIs) for prompt gamma (PG) origins at a coarse level; the second tier utilizes an enhancement model to restore true PG origins by applying additional attention within ROIs. Clinical data from a prostate cancer patient treated with proton therapy were extracted, including planning computed tomography (CT) and treatment plans. Using the Geant4 Monte-Carlo toolkit, we simulated PG emissions during proton pencil beam delivery (energy: 157.6 MeV - 198.7 MeV; beam angle: 90ยฐ) on the patientโs anatomy, comprising 25 distinct materials. Beam spots were combined to achieve a dose level >100 MU per sample, generating a dataset of 174 samples (80% for training; 20% for testing). PGI reconstruction was performed using a filtered back-projection algorithm, enhanced by the DL model. The efficacy of the DL-enhanced PGI was assessed using proton range errors (ฮR).
Results: DL-enhanced PGI showed clear and accurate proton ranges along beam direction: ฮR of beam starting point is 2.7mm ยฑ 2.6mm, and ฮR of beam ending point is 2.0mm ยฑ 1.7mm.
Conclusion: The proposed attention-driven DL method effectively restores PG origins and determines proton beam ranges within patient anatomy from single CC measurements, significantly improving PGIโs clinical utility for real-time proton range verification.