PROTOTWIN-PET: A Deep Learning and GPU-Based Workflow for Dose Verification in Proton Therapy With PET

Abstract

In proton therapy (PT), accurate dose delivery verification is critical for detecting treatment plan deviations. This can be achieved by imaging activated positron emitters with a positron emission tomography (PET) acquisition and converting the data into a delivered dose image. This work presents PROTOTWIN-PET (PROTOn therapy digital TWIN models for dose verification with PET), a patient-specific, deep learning (DL) and GPU-based workflow for 3-D dose verification. The proposed workflow generates a dataset of simulated, realistic 3-D PET and dose pairs that reflect possible clinical deviations in patient positioning and physical parameters. Using this dataset, a DL model is trained to estimate the delivered dose from the PET image, incorporating a deviation-predicting branch (DPB) to estimate patient positioning deviations. PROTOTWIN-PET is demonstrated on a two-field oropharyngeal cancer treatment plan, estimating the delivered dose in milliseconds with an average mean relative error of 0.6% and near-perfect gamma passing rates (3 mm, 3%). Positioning deviations are estimated on average within a tenth of a millimeter and degree. PROTOTWIN-PET can be implemented within the one-day interval between the plan CT acquisition and the first treatment session, potentially enabling timely treatment plan adjustments and maximizing the precision of PT. PROTOTWIN-PET is available at github.com/pcabrales/prototwin-pet.git.