Speaker
Description
Range uncertainty remains a major limitation in proton therapy because treatment planning depends on CT‐to‐stopping‐power conversion, which introduces an intrinsic uncertainty of approximately 3.5%. Discrete range modulation (DRM) proton radiography (pRG) provides a projection-based estimate of the water equivalent path length (WEPL) by delivering a sequence of discrete proton energies and analyzing the energy-resolved dose curve (EDC) measured at each detector pixel. In anatomically complex cases, however, multiple Coulomb scattering (MCS) broadens the lateral response and can bias WEPL estimates.
In this study, we developed a two-stage deep-learning framework to improve WEPL estimation accuracy and enable practical DRM pRG for range verification in head-and-neck cases. Proton energies from 70 to 230 MeV were simulated in 2 MeV increments. A virtual 2D detector (250×250 pixels over 250×250 mm²) recorded energy-resolved dose curves (EDCs) for each layer. In Stage 1, a three-dimensional convolutional neural network (3D CNN) predicted the WEPL at the central pixel from a local input tensor constructed from the full EDC stack across energy layers and a spatial neighborhood around that pixel; neighborhood (patch) sizes from 5×5 to 13×13 were evaluated. Training samples were stratified by WEPL bins to mitigate class imbalance and avoid over-representation of low-WEPL regions. In Stage 2, a refinement network incorporated the Stage-1 WEPL output together with a co-registered DRR image to enhance local spatial resolution.
The result shows that Stage-1 network achieved a mean WEPL RMSE of 2.87 mm (1.27%) and a mean 95th-percentile relative error of 2.57%. The Stage-2 model achieved a mean WEPL RMSE of 1.28 mm (0.58%) and a mean MAE of 0.93 mm (0.42%), with a mean 95th-percentile relative error of 1.16%. These results indicate that the proposed two-stage approach can reduce MCS-related WEPL errors and improve the feasibility of DRM pRG for range-verification workflows in head-and-neck proton therapy.