ESTRO38 Congress Report

Physics

1. MRI-only proton therapy treatment planning with synthetic CT images generated using deep learning (E38-2116) Ana María Barragán Montero 1,2 , Samaneh Kazemifar 2 , Kevin Souris 1 , Robert Timmerman 3 , Steve Jiang 2 , Xavier Geets 1 , Edmond Sterpin 1 , Amir Owrangi 2 . 1 UCLouvain- Institute of Experimental and Clinical Research, Molecular Imaging, Radiotherapy and Oncology MIRO, Brussels, Belgium, 2 UT Southwestern Medical Center, Medical Artificial Intelligence and Automation MAIA, Dallas, USA. 3 UT Southwestern Medical Center, Radiation Oncology, Dallas, USA.

Context of the study Magnetic resonance imaging (MRI) is often used in radiation therapy treatments to accurately contour the target volume and organs at risk (OARs), due to the superior soft tissue contrast in comparison with computed tomography (CT) images. The use of MRI images is especially crucial in treatment sites located in the abdomen and brain, where the tumour volume is mainly surrounded by soft tissue. However, CT images are still required in order to retrieve information needed for dose calculation (i.e., electron density values for conventional radiotherapy with photons and stopping powers for ion therapy). Therefore, the current treatment planning workflow for these sites relies on the contouring of the target and OARs on MRI and the posterior transfer of contours to CT via image registration. However, MRI-CT co-registration introduces systematic geometrical uncertainties up to 2-3 mm, which shift high dose regions away from the target and lead to a geometric miss that compromises tumour control. This problem has recently led to the concept of MRI-only based treatment planning, where pseudo or synthetic CT images for dose calculation are generated directly from the MRI scan. In addition, treatment planning based on MRI-only would reduce radiation dose, patient time, and hospital resources. MRI-only treatment planning is then a very attractive concept that is gaining popularity. However, the accurate generation of Hounsfield unit (HU) maps fromMRI images is not straight-forward. The rise of deep learning methods and the promising results obtained in medical imaging applications provides a good environment to apply this kind of methods to the synthetic CT generation problem. In this study, we evaluate the dosimetric accuracy of synthetic CT images generated with a deep learning method based on generative adversarial networks (GANs), for treatment planning of scanned proton therapy. Overview of abstract MRI-only treatment planning is becoming popular due to increased soft-tissue contrast and the reduction of radiation dose, among other advantages. Since CT images are still needed for dose calculation, different methods have been proposed to convert MRI to synthetic CT (sCT) images. Specifically for proton therapy, sCT must have extremely good quality, because small HU errors could greatly impact the dose distribution. This work explores the application of deep learning to this problem, since it could potentially improve the quality of the generated sCT images in comparison with the existing methods, given its ability to capture complex relationships between images.

What were the three main findings of your research? 1. High quality images: Our deep learning model based on generative adversarial networks (GANs) was able to learn how to accurately transform MRI to synthetic CT images, with a mean absolute error over all test patients as low as 47.2 ± 11.0 HU. 2. Easy implementation: The model generated the sCT images in less than 10 s, without requiring any human intervention and using a single MRI sequence (T1- weighted). 3. High dosimetric accuracy: The dosimetric evaluation showed that the dose differences were mostly below 1 to 2% of the dose prescription, which outperforms the current existing models for synthetic CT generation. What impact could your research have? So far, themodels presented in the literature requiredmanual pre- or post- processing of the synthetic CT images, such as manual insertion of air-cavities or bone segmentation, in order to minimize proton range differences and ensure a reasonable dosimetric accuracy. In contrast, our model removes all human intervention and achieves a high accuracy, even in cases with complex air-bone interfaces close to nasal cavities. This, together with the superior speed for the sCT generation (less than 10s), brings closer the clinical implementation of MRI-only treatment planning. Thus, potentially reducing the uncertainties due to MRI to CT contour propagation inherent to the current treatment planning workflow, which may improve clinical outcomes in those tumours located in soft-tissue regions. Is this research indicative of a bigger trend in oncology? Indeed, the application of deep learning methods in the domain of medical imaging is becoming more and more trendy, motivated by the excellent results obtained so far. This work is just another example of this trend, which justifies the investigation of deep learning for new applications, and eventually, the clinical implementation of some of them in the near future. However, before bringing these models to the clinical environment, the scientific communitymust carefully validate themwith databases that are representative of the targeted patient population, and build quality assurance tools to ensure its correct performance for each single case and/or patient.

PHYSICS | Congress report

14