Click here to

Session: Treatment Planning and Delivery in Particle Therapy [Return to Session]

Deep Learning-Based Prompt-Gamma Imaging for Range Verification in Proton Radiotherapy

Z Jiang1*, J Polf2, L Ren3, (1) Duke University, Durham, NC, (2) University of Maryland School of Medicine, Baltimore, MD, (3) University of Maryland, Baltimore, MD

Presentations

SU-J-206-3 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]

Room 206

Purpose: Range uncertainty is a major bottleneck in proton therapy, limiting its delivery accuracy. Compton camera (CC)-based prompt-gamma (PG) imaging is a promising technique for range verification. However, back-projection-based PG reconstruction suffers from severe distortions due to limited-view measurement by CC, significantly limiting its clinical utility. We proposed a deep learning-based method to generate accurate PG images to achieve precise proton range verification.

Methods: PG imaging has two challenges: (1) Signals are severely stretched towards CC and almost indistinguishable in the back-projected PG images, and (2) PG signals emitted along a pencil-beam path take up an extremely low portion of 3D space, presenting an attention challenge for deep-learning. To solve these issues, we proposed a two-tier scheme: first, a localization network is trained to define a region-of-interest (ROI) in the distorted PG image that contains the proton pencil beam; second, an enhancement network is trained to restore the true PG signals within the ROI. In this study, 21 proton pencil-beams with energies from 75MeV-125MeV and a clinical dose rate (20kMU/min) were simulated in a tissue-equivalent phantom. PG detection with a CC was simulated using the Monte-Carlo-Plus-Detector-Effects (MCDE) model. Images were reconstructed using back-projection and were then enhanced by the proposed method. We left out 80Mev, 100MeV, and 120MeV beams for testing, and used the other 18 beams for training and validation. Results are evaluated qualitatively and quantitatively using range errors (ΔR) and relative mean absolute errors (RMAE).

Results: PG images were significantly improved matching closely to the ground-truth. ΔRs were <1 pixel, <1 pixel, 1 pixel for 80MeV, 100MeV, and 120MeV beams, respectively (pixel size of 2.0 mm). RMAEs were 1.56%, 1.25%, 1.24% for 80MeV, 100MeV, and 120MeV beams, respectively.

Conclusion: Our study demonstrated the feasibility of deep-learning to generate highly accurate PG images to provide precise proton range verification.

Keywords

Gamma Cameras, Dosimetry, Protons

Taxonomy

TH- External Beam- Particle/high LET therapy: Range verification (in vivo/phantom): prompt gamma/PET

Contact Email

Share: