ESTRO 2024 - Abstract Book

S4495

Physics - Machine learning models and clinical applications

ESTRO 2024

References:

[1] Engelsman, M., Schwarz, M., & Dong, L. (2013). Physics Controversies in Proton Therapy. Seminars in Radiation Oncology, 23(2), 88–96. https://doi.org/10.1016/j.semradonc.2012.11.003

[2] Yuxi Li, “Deep Reinforcement Learning: An Overview”, CoRR, 2017, http://arxiv.org/abs/1701.07274

1753

Digital Poster

Deep learning for in vivo EPID dosimetry classification: Relating gamma analysis and DVH metrics

Femke Vaassen, Sebastiaan Nijsten, Richard Canters, Frank Verhaegen, Cecile Wolfs

GROW School for Oncology and Reproduction, Maastricht University Medical Centre+, Department of Radiation Oncology (Maastro), Maastricht, Netherlands

Purpose/Objective:

For lung cancer patients, anatomical changes during radiotherapy treatment can result in significant differences between planned and delivered dose [1]. This makes it essential to perform in-treatment, (i.e. in vivo) verification of delivered dose to ensure treatment quality for the patient [2]. The electronic portal imaging device (EPID) has previously shown to be able to accurately and efficiently perform in vivo dose verification [3,4]. Measured EPID doses can be compared to predicted/planned dose distributions using gamma (γ) analysis. The result of this comparison is usually summarized in one or a few numbers, providing fast decision making based on fixed thresholds using gamma pass rate. However, this leads to loss of 2D and/or 3D dosimetric and time-resolved information. In this study, a deep learning classification model is trained on 2D γ-maps, investigating the possibility to relate 2D γ analysis to clinically relevant dose volume histogram (DVH) metrics for in vivo treatment verification using EPID dosimetry. The aim of the model was to classify γ-maps into two categories: ‘error’ and ‘no error’.

Material/Methods:

EPID measurements during 3086 fractions of 363 lung cancer patients (380 unique treatment plans) were converted to portal dose images (PDIs) using in-house developed software [5]. Predicted transit PDIs were created based on the planning CT. Measured and predicted PDIs were compared per beam using γ-analysis with 5%/3mm criteria, resulting in 5844 γ maps. The γ-maps were cropped using a 10% low dose threshold, normalized using min/max normalization and resized to 128x128 pixels. The ground-truth classification (error vs. no error) was based on DVH metrics derived from 3D Monte Carlo dose recalculations on daily cone-beam CT (CBCT) images, including contour propagation from planning CT to CBCT. Different ground-truth classification methods were implemented, based on exceeding the clinical constraints