Deep Learning-Based 3D Proton-Acoustic Imaging Using a Matrix Ultrasound Array for Dose Verification in Prostate Proton Therapy
Z Jiang1*, L Sun2, W Yao3, L Xiang4, L Ren5, (1) Duke University, Durham, NC, (2) University of California Irvine, Irvine, CA, (3) University of Maryland School of Medicine, Baltimore, MD, (4) University of California, Irvine, Irvine, CA, (5) University of Maryland, Baltimore, MD
Presentations
TU-D930-IePD-F2-2 (Tuesday, 7/12/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 2
Purpose: Dose delivery uncertainty caused by motion or range uncertainty is a major concern in proton therapy. A promising technique, proton-acoustic imaging, has been proposed to verify dose by measuring radiation-induced pressures. However, its dosimetry accuracy is severely limited by the poor image quality due to limited view. We aim to address this bottleneck by developing a deep-learning-based method to reconstruct high-quality 3D dosimetry using a matrix ultrasound array for dose verification in prostate proton therapy.
Methods: The proposed method consists of two models, which were cascaded and jointly trained: the first model learned to reconstruct accurate pressure maps from the limited-view ultrasound acquisition; the second model learned to derive dose maps from pressure maps. To evaluate model performance, we used81 prostate patients’ proton therapy treatment plans with dose calculated using RayStation and normalized to the maximum dose. The proton-acoustic propagation was simulated using the open-source k-wave package. A matrix ultrasound array containing 64×64 sensors was simulated near the perineum to receive acoustic signals during dose delivery. The model was trained using 69 patients and tested using 12 other patients. Results were evaluated both qualitatively and quantitatively using root-mean-squared-error (RMSE), gamma-index (GI), and isodose line Dice.
Results: The proposed method considerably improved the limited-view proton-acoustic image quality, demonstrating clear and accurate edges and structures in the reconstructed pressure maps and high agreement between the reconstructed and ground-truth dose maps. Quantitatively, the pressure map achieved an RMSE of 0.067. The dose map reached an RMSE of 0.049, GI (3mm,3%) of 92.10%, and 90% isodose line Dice of 0.914.
Conclusion: To our knowledge, this is the first time high-quality quantitative 3D dosimetry is demonstrated in proton-acoustic imaging. Using a matrix array can potentially enable real-time 3D dose verification during proton therapy to significantly improve its precision and outcomes.
Keywords
Not Applicable / None Entered.
Taxonomy
Contact Email