AI-Driven Radiacoustic Imaging for Proton Therapy Monitoring
Abstract
Purpose
Real-time dose verification remains a critical unmet need in radiation therapy. We developed physics-informed AI models for radiacoustic imaging (RAI) to enable quantitative, real-time in vivo dose monitoring for proton therapy. Our approach addresses two big challenges: electromagnetic interference and limited-view problem through AI-driven denoising and image reconstruction models, paving the way for adaptive proton therapy with immediate treatment feedback.
Methods
At OUHSC, proton-induced RA signals were collected from water and soft tissue phantoms, with each spot irradiated >500 times. Individual frame signals exhibit strong electromagnetic interference, achieving reasonable SNR (>5 dB) only after averaging hundreds of frames. For clinical applicability, we developed a U-Net denoising model that predicts high-quality reference signals (equivalent to >500-frame averages) from only 10 frames, enabling real-time dose monitoring. For dose reconstruction, we developed a physics-informed neural network (PINN) that reconstructs dose maps from limited-aperture (5cm × 5cm) transducer data. The PINN incorporates acoustic wave physics and digital twins of the radiation delivery and detection systems, enabling accurate and physically-consistent reconstructions from limited-view data. Cadaver liver irradiations with 154 MeV protons validated this pipeline: 10-frame averaged sinograms were denoised and then reconstructed via PINN, and the outcomes were compared with the treatment plans.
Results
Across three cadaver datasets, denoising achieved 91%±1% correlation with reference signals while effectively eliminating electromagnetic interference. Dose maps reconstructed from these denoised sinograms demonstrated strong agreement with planned dose distributions, achieving 90%±3% gamma pass rates using 5mm/5% criteria and 10% dose threshold, validating the clinical feasibility of our approach.
Conclusion
Proton-RAI is well-positioned for patient studies, as demonstrated through cadaver experiments integrating sophisticated data acquisition with AI-driven denoising and reconstruction models. The in vivo monitoring capability enables adaptive proton therapy, where RAI-reconstructed dose maps provide real-time feedback for beam delivery adjustments, enhancing treatment precision and outcomes.