ESTRO 2023 - Abstract Book

S1869

Digital Posters

ESTRO 2023

Conclusion Developing radiobiological models of response to radioimmunotherapy is important to assist in the design of optimal administration schedules and dosages in radioimmunotherapy. Approaches like the one presented in this work constitute a stepping stone towards a better understanding of the response of tumors to radioimmunotherapy, even if these preliminary pre-clinical results must be considered with care.

PO-2086 Machine learning to predict local failure in melanoma brain metastases treated with radiosurgery

M. Shanker 1 , P. Ramachandran 2 , R. Motley 2 , M. Huo 2 , M. Foote 2 , M. Pinkham 2

1 The University of Queensland, Radiation Oncology, Brisbane, Australia; 2 Princess Alexandra Hospital, Radiation Oncology, Brisbane, Australia Purpose or Objective Radiomics has the potential to revolutionize clinical decision making in the management of melanoma brain metastases (MBM) treated with stereotactic radiosurgery (SRS). Radiomic features extracted from baseline magnetic resonance (MR) imaging can be integrated into clinicoradiological parameters to predict long term outcomes in order to tailor multimodal treatment strategies, permit individualized surveillance imaging frequency and facilitate early changes to therapy following SRS. We present the first Australian institutional data assessing the predictive accuracy of a baseline-MRI radiomics model in MBM patients to predict local failure following SRS. Materials and Methods Patients receiving single-fraction Cobalt-based SRS for MBM at a single Australian institution were analysed. Predictive patient and treatment characteristics, including the type and timing of systemic therapy were collected. Progression of disease (PD) outcomes were defined either histologically or according to RANO-BM criteria. 108 radiomic features were extracted from T1-weighted Gadolinium contrast-enhanced MPRAGE MRI sequences using in-house software developed in MATLAB. Highly dependent radiomic features were selected using an analysis of variance (ANOVA). Four models; logistic regression, random forest, SVC, and decision tree models were generated to predict local progression as a binary classifier. A multivariate model was additionally developed, integrating radiomic features with baseline lesion volume, immunotherapy use and SRS dose. The models’ accuracy and precision score were computed to assess the power of the model to predict progression of disease. Results 101 MBM patients were treated with SRS. The median duration of follow-up was 29.2 months (IQR 19.7-39.8). Median dosage was 20Gy (IQR 18-20). The median volume and diameter of the lesion at baseline were 0.24cc (IQR 0.06-1.02) and 7.7mm (IQR 4.8-12.2), respectively. 53% of patients were BRAF mutants, and 65% of those patients had failed BRAF inhibitors prior to SRS. 34.4% of patients took BRAF inhibitors concurrently (48 hours), and 77.0% received immunotherapy concurrently (4 weeks pre to 4 weeks post-SRS). 349 MBM lesions were included in the radiomics model. Utilizing baseline imaging alone, all four models were able to predict long term PD following SRS with an average 85% prediction accuracy and an 86% precision score. Homogeneity, Correlation, Joint Entropy, Cluster Prominence, and Difference Entropy were the prominent features among all the extracted features and have the potential to predict PD. Conclusion A baseline-MRI radiomics model in MBM patients can predict local failure following SRS to a high degree of accuracy. Additional integration of radiomics models utilizing multiparametric imaging combined with patient and treatment characteristics will optimize the use of radiomic tools into the clinic. 1 KU Leuven, Laboratory of experimental radiotherapy, Leuven, Belgium; 2 UZ Leuven, Department of radiation oncology, Leuven, Belgium; 3 KU Leuven, Department of radiation oncology, Leuven, Belgium; 4 AZ Turnhout, Department of radiation oncology, Turnhout, Belgium; 5 Université catholique de Louvain, Institut de Recherche Expérimentale et Clinique, Center of Molecular Imaging, Radiotherapy and Oncology (MIRO), Brussels, Belgium Purpose or Objective Approximately 30% of patients treated for locally advanced oesophageal cancer develop pulmonary and/or cardiac complications after neoadjuvant radiochemotherapy and surgical resection. This study aims to expose risk factors for these complications and to investigate the potential of features extracted from 3D dose distributions (dosiomics) in normal tissue complication probability (NTCP) modelling. Materials and Methods Patients treated with trimodality therapy in UZ Leuven from 2011 to 2021 were included. Model building (122 patients, 2011-2018, 3DCRT or IMRT) and validation (47 patients, 2018-2021, IMRT) sets were defined based on treatment year. The prescription dose was 45 Gy in fractions of 1.8 Gy. NTCP models were constructed and validated for pulmonary and cardiac postoperative complication endpoints. Clinical factors (age, BMI, tumour histology etc.) and DVH features (relative VxGy and mean dose for left lung, right lung, lungs combined, heart and left ventricle) were combined with dosiomics features calculated from the 3D dose distributions in the lungs and the heart. These dosiomics features include first- and second-order statistics and matrix-based texture features derived from the organ dose maps. Dosiomics extraction was implemented using the PyRadiomics library. A two-step multivariate logistic regression analysis was performed. In the first step, a repeated 5-fold cross-validation PO-2087 NTCP modelling with dosiomics features for postoperative complications in oesophageal cancer R. Duwaerts 1,4 , G. Defraene 1 , P. Populaire 1,2 , E. Sterpin 1,5 , K. Haustermans 1,3