ESTRO 2025 - Abstract Book

S3449

Physics - Machine learning models and clinical applications

ESTRO 2025

References: 1 Gooding MJ, Aluwini S, Guerrero Urbano T, McQuinlan Y, Om D, Staal FHE, Perennec T, Azzarouali S, Cardenas CE, Carver A, Korreman SS, Bibault JE. Fully automated radiotherapy treatment planning: A scan to plan challenge. Radiother Oncol. 2024 Aug 31:110513. doi: 10.1016/j.radonc.2024.110513. Epub ahead of print. PMID: 39222848. 2 Cardenas CE, Cardan RA, Harms J, Simiele E, Popple RA. Knowledge-based planning, multicriteria optimization, and plan scorecards: A winning combination. Radiother Oncol. 2024 Oct 26:110598. doi: 10.1016/j.radonc.2024.110598. Epub ahead of print. PMID: 39490417.

4181

Digital Poster MIND (Magnetic resonance Imaging Neural Dose calculator): Monte Carlo Level MRI-Only Proton Dose Calculation via Neural Networks Muheng Li 1,2 , Carla Winterhalter 1 , Xia Li 1,3 , Sairos Safai 1 , Antony John Lomax 1,2 , Ye Zhang 1 1 Center for Proton Therapy, Paul Scherrer Institut, Villigen, Switzerland. 2 Department of Physics, ETH Zürich, Zürich, Switzerland. 3 Department of Computer Science, ETH Zürich, Zürich, Switzerland Purpose/Objective: Despite MRI's superior soft-tissue contrast, current proton therapy workflows still rely on CT imaging for dose calculations. While synthetic CTs can be generated from MR images [1], such pipeline introduces additional computational overhead. This study presents MIND, a deep learning-based dose engine enabling proton dose calculation from MR images directly, aiming to streamline the treatment planning workflow efficiently while maintaining Monte Carlo (MC)-level dose accuracy. Material/Methods: The study utilized paired MR-CT scans from 39 brain tumour patients (29/3/7 for training/validation/testing). We developed a deep learning framework (Figure 1a) including convolutional neural networks with various sequence models (LSTM [2], Transformer [3], Mamba [4], and xLSTM [5] architectures) to predict individual proton pencil beam dose distributions directly from MR image. For training, 2000 random beam configurations were generated for each patient, varying in gantry angle (0-180°), couch angle (0-180°), and energy (70-181 MeV), and FRED [6] MC dose distributions were pre-calculated with high statistics (1 million protons/beam). The framework processes beam-eye-view (BEV) volume patches sampled at 2mm isotropic resolution and predicts corresponding dose distributions. During model testing, the MIND derived dose distributions were directly compared against MC results. Additionally, for comparison, the corresponding model trained by CT images was evaluated. The evaluation was performed for both beam-wise dose predictions (Figure 1b) and full treatment plan evaluation (Figure 2). Results: Accurate predictions of beam-wise dose distributions were obtained from all MIND modes, closely matching to corresponding CT-based results (Figure 1b). For full plan evaluation, comparative analysis of different sequence models revealed that xLSTM architecture achieved the overall highest performance for both MR and CT-based models (Figure 2a). The best MIND mode achieved gamma pass rates of 99.57±0.44% (1mm/1%) and mean percentage dose errors (MPDE) of 0.60±0.25% within patient bodies and 1.73±0.83% in regions receiving more than 50% of the prescribed dose. Comparison of dose distributions and DVHs for representative cases (Figure 2b) confirmed that MIND calculations achieved visually similar plans. Notably, MIND requires only 3ms for each spot dose prediction, compared to 2s needed by FRED MC simulation with 10 7 particle histories.