ESTRO 2024 - Abstract Book

S5027

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

J, Wohlfahrt P, Lomax T. Roadmap: proton therapy physics and biology. Phys Med Biol. 2021 Feb 26;66(5):10.1088/1361-6560/abcd16. doi: 10.1088/1361-6560/abcd16. PMID: 33227715; PMCID: PMC9275016.

1182

Poster Discussion

Deep learning-based multi-modal approach for predicting brain radionecrosis after proton therapy

Sithin Thulasi Seetha 1,2 , Giulia Fontana 2 , Alessia Bazani 3 , Giulia Riva 4 , Silvia Molinelli 3 , Christina Amanda Goodyear 5 , Lucia Pia Ciccone 4 , Alberto Iannalfi 4 , Ester Orlandi 4 1 University of Pavia, Department of Clinical, Surgical, Diagnostic and Pediatric Sciences, Pavia, Italy. 2 National Center for Oncological Hadrontherapy (CNAO), Clinical Department, Pavia, Italy. 3 National Center for Oncological Hadrontherapy (CNAO), Medical Physics Unit, Clinical Department, Pavia, Italy. 4 National Center for Oncological Hadrontherapy (CNAO), Radiation Oncology Unit, Clinical Department, Pavia, Italy. 5 University of Naples Federico II, Department of Advanced Biomedical Sciences, Naples, Italy

Purpose/Objective:

The development of brain radionecrosis (BRN) was reported as a detrimental side effect of proton beam radiotherapy (PBT) for the treatment of malignancies in the brain, head and neck, and skull-base regions [1, 2]. Early assessment of healthy tissues vulnerable to necrosis due to radiation exposure could facilitate treatment optimization. To this account, we propose a deep learning approach to predict the voxel-wise likelihood of BRN using multi-modal (CT, MR) brain images, and dose maps.

Material/Methods:

We gathered retrospective data from 41 meningioma and hemangiopericytoma patients treated with PBT between February 2014 and November 2021. Data includes computed dose maps and multimodal pre-treatment simulation images such as planning CT, post-contrast T1-weighted volumetric interpolated breath-hold examination (T1v), and fat-saturated (fs) T2-weighted turbo spin echo (T2w), all acquired with customized immobilization systems. The imaging acquisition protocols are summarized in Figure 1. All the images were rigidly co-registered with the planning CT using RayStation v.11B (RaySearch, Stockholm, Sweden) embedded registration software. After the treatment, follow-up MR images were acquired every three months for the first year, every three-six months for the subsequent two years, and annually thereafter. The incidence of BRN was radiologically assessed on MRI follow-up exams, and all the necrotic regions were segmented by a skilled radiotherapist after rigid registration with planning CT.