ESTRO 2025 - Abstract Book

S3440

Physics - Machine learning models and clinical applications

ESTRO 2025

should be taken when relying only on point predictions from ML models. Predictions providing low GPRs are less reliable than those yielding high GPRs. Supported by: GR-2019-12370739.

Keywords: uncertainty quantification, multicentric, PSQA

References: [1] Lambri N, Hernandez V, Sáez J, et al. Multicentric evaluation of a machine learning model to streamline the radiotherapy patient specific quality assurance process. Phys Med. 2023;110:102593. doi:10.1016/j.ejmp.2023.102593 [2] Claessens M, De Kerf G, Vanreusel V, et al. Multi-institutional generalizability of a plan complexity machine learning model for predicting pre-treatment quality assurance results in radiotherapy. Phys Imaging Radiat Oncol. 2023;29:100525. Published 2023 Dec 19. doi:10.1016/j.phro.2023.100525

3978

Digital Poster A Deep Learning model for fast and accurate Oropharynx patients’ pretreatment positioning prediction in proton therapy Hooman Bahrdo, Gabriel Guterres Marmitt, Femke Oosterhof, Johannes Langendijk, Stefan Both Radiation Oncology, University Medical Center Groningen (UMCG), Groningen, Netherlands Purpose/Objective: Pretreatment patient positioning is key for accurate intensity-modulated proton therapy (IMPT) treatment delivery. Currently ,in proton therapy (PT), radiation therapists perform Cone-Beam-Computed-Tomography-based semi automated patient positioning in up to 15minutes (min). Positioning time and accuracy may vary with therapist skills and patient anatomy changes. Therefore, we aim to develop a fast, accurate, and precise self-supervised deep learning (DL) model to help ensure optimal automated dose-guided patient positioning (DGPP) in PT. Material/Methods: This study included 53 oropharyngeal-cancer patients (OCP) treated with IMPT and receiving in average 6 weekly treatment verification CTs (vCT) per institutional clinical protocol. A treatment setup uncertainty of 3.00millimeters (mm) is employed in our clinical practice for OCPs. The optimal relative positions of vCTs to planning CTs (pCTs) were calculated using a dose-guided gradient descent algorithm [1]. Training and validation vCT data were augmented by shifting the treatment isocenter up to 15.00mm 80-times to mimic patient setup variations, while the 6-patient test set was augmented 10-times with offsets up to 7.00mm as encountered in clinical practice. Dose distributions generated by Moqui[2] on vCTs and pCTs served as inputs for a Dual-CNN model, which used convolutional layers with Leaky-ReLU, batch normalization, fully-connected layers, and dropouts to predict treatment isocenter shift vectors (x,y,z). Absolute Error (AE), the distance between optimal and predicted isocenter positions, was used to evaluate accuracy and precision. Results: The model's performance on the test set (cohort) is summarized in Figures 1 and 2. Figure1 shows the AE distribution of the cohort, with mean error of 1.26mm (95% CI: 1.14 – 1.36mm), all within the 3.00mm treatment robustness threshold. 69.00% of prediction errors are within 1.50mm and 31.00% fall between 1.50 – 3.00mm, with only 1.52% exceed 2.50mm, demonstrating model accuracy for new oropharyngeal patients. Figure2a displays AE distributions per patients/weekly-vCTs case, and Figure2b shows their Standard Deviation (STD). Most prediction inconsistencies of the cases are under 1.50mm, and the average STD is 0.34mm (cohort STD: 0.57mm), demonstrating precision and consistency despite variations in initial treatment isocenter positions. Predictions are computationally efficient, averaging 0.035±0.001seconds-per-case.