ESTRO 2024 - Abstract Book

S4971

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Oncology, Zurich, Switzerland. 5 University Hospital Magdeburg, Department of Radiation Oncology, Magdeburg, Germany. 6 University Hospital Würzburg, Julius-Maximilians-University, Department of Radiation Oncology, Würzburg, Germany. 7 Klinikum rechts der Isar, Technical University of Munich, Department of Neurosurger, Munich, Germany. 8 Heidelberg University Hospita, Department of Radiation Oncology, Heidelberg, Germany. 9 Kantonsspital Aarau, Radiation Oncology Center KSA-KSB, Aarau, Switzerland. 10 General Hospital Fulda, Department of Radiation Oncology, Fulda, Germany. 11 University Medical Center Schleswig Holstei, Department of Radiation Oncology, Kiel, Germany. 12 University of Freiburg - Medical Center, Department of Radiation Oncology, Freiburg, Germany. 13 Saphir Radiosurgery Center Frankfurt and Northern German, Department of Neurosurgery, Kiel, Germany. 14 German Cancer Consortium (DKTK), Partner Site Freiburg, Freiburg, Germany. 15 German Cancer Consortium (DKTK), Partner Site Munich, Munich, Germany. 16 Helmholtz Center Munich, Institute of Radiation Medicine (IRM), Department of Radiation Sciences (DRS), Munich, Germany. 17 Klinikum rechts der Isar, Technical University of Munich, Institute of Radiation Medicine (IRM), Department of Radiation Sciences (DRS), Munich, Germany In the clinical management of cancer, time-to-event prediction plays an important role in accurately modeling the overall survival of patients and the time to progression. For patients with brain metastases (BMs), the improved prognosis of the recurrence of the metastasis (i.e., local failure, LF) could be beneficial in determining the optimal radiotherapy dose concept. In this work, we use deep convolutional neural networks (CNNs) for time-to-event prediction of metastasis recurrence based on pre-treatment status for patients who received resection and post-operative stereotactic radiotherapy. In an end-to-end fashion, our model combined imaging features from MRI and biomarkers as a multi-modal approach Purpose/Objective: Our dataset consisted of 352 patients collected from 7 different centers within the retrospective AURORA study of the “Radiosurgery and Stereotactic Radiotherapy Working Group” of the German Society for Radiation Oncology. All patients received surgical resection and post-operative stereotactic radiotherapy to the resection cavity. For each patient, pre-treatment MRI sequences of T1-CE and FLAIR are available alongside clinical characteristics such as age , Karnofsky index , primary cancer type , metastasis locatio n, and binary indicators of chemotherapy and immunotherapy received during treatment of the primary cancer. As the right-censored survival endpoint, time-to-LF of the largest metastasis was used. Inspired by Mobadersany et al. [4], our model consisted of a convolutional neural network to extract image features and project them into hazard ratios. As the backbone, we employed a ResNet18 [3] model adapted to work in 3-D space. For a more focused input view, we cropped the images to the bounding around the edema tissue of the gross tumor volume (GTV) of the largest metastasis and filtered out any other lesion. In order to detect lesions for filtering and cropping, we use an auto-segmentation model published by Buchner et al. [1]. Material/Methods: