ESTRO 2025 - Abstract Book

S3430

Physics - Machine learning models and clinical applications

ESTRO 2025

3625

Digital Poster Investigating the feasibility of deep-learning-based fast MRI-only proton range verification Liheng Tian, Armin Lühr Physics, TU-Dortmund, Dortmund, Germany Purpose/Objective: Proton therapy dose is highly susceptible to anatomical changes. Pre-treatment magnetic resonance imaging (MRI) could provide monitoring of changes of anatomy and particularly the target volume. However, conventionally it is insufficient to calculate proton dose. In this project, we investigated the possibility to detect proton range changes using fast MRI-only proton dose calculation taking advantage of state-of-the-art deep learning (DL) methods. Material/Methods: For sCT generation, the DL model pix2pix [1] was trained using an open-access dataset of registered MRI and CT images of 120 pelvis patients [3] and applied to generate an sCT for each patient. For proton dose prediction, first, Monte Carlo (MC) simulations of 20k pencil beams (PB) with random energies between 70-220 MeV were performed with TOPAS on patient CT using random beam angles and positions. Then, these PBs were used to train a DL model, DoTA [2], to predict proton dose on CT. For evaluation, additional 1500 random PBs were simulated with MC on, both, CT and corresponding sCT. For these PBs, dose prediction was performed applying the trained DoTA on corresponding CT/sCT. The performance of DoTA was measured by comparing DL and MC proton range (80% distal dose fall-off) on the same image modality. The proton range difference induced by image mismatch was measured by comparing MC dose based on CT and MRI. Figure 1 shows an example evaluation. Results: The difference (mean±standard deviation) between CT numbers of CT and generated sCT in the beam path was - 0.5±87.5 HU and the mean absolute error was 35.6 HU. This image mismatch induced a proton range difference of - 0.2±3.0 mm (Fig. 2B, orange). The deep learning dose prediction model DoTA trained only on the CT dataset performed equally well on CT and sCT, the latter being derived from MRI. The difference between MC-simulated and DL-predicted proton ranges were 0.2±1.4 mm and 0.2±1.3 mm for CT and MRI-derived sCT, respectively, i.e., not significantly different (Figure 2A). Finally, the difference between MRI-only DL predicted dose and MC dose on CT (0.0±3.0 mm) was compatible to the difference induced by image mismatch (-0.2±3.0 mm cf. above), which can be caused by, e.g., anatomical changes between CT and MRI acquisition. Conclusion: The proton range error caused by DL dose prediction is smaller than that caused by CT-MRI mismatch. Therefore, fast proton range verification directly on pre-treatment MRI appears possible using DL MRI-only dose prediction. References: [1] L. Tian, A. Lühr, Proton range uncertainty caused by synthetic computed tomography generated with deep learning from pelvic magnetic resonance imaging, Acta Oncol. 62 (2023) 1461 – 1469. https://doi.org/10.1080/0284186X.2023.2256967. [2] O. Pastor-Serrano, Z. Perkó, Millisecond speed deep learning based proton dose calculation with Monte Carlo accuracy, Phys. Med. Biol. 67 (2022) 105006. https://doi.org/10.1088/1361-6560/ac692e. [3] A. Thummerer, E. Van Der Bijl, A. Galapon, J.J.C. Verhoeff, J.A. Langendijk, S. Both, C. (Nico) A.T. Van Den Berg, M. Maspero, SynthRAD2023 Grand Challenge dataset: Generating synthetic CT for radiotherapy, Med. Phys. 50 (2023) 4664 – 4674. https://doi.org/10.1002/mp.16529. Keywords: Proton range, deep learning