S2487
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Conclusion: Our study leverages multi-sequence MRI to non- invasively improve case stratification versus historical image-based methods. Preliminary results suggest perfusion MRI-derived imaging biomarkers can assist TP/RN classification. Expanding upon this work, we aim to incorporate novel perfusion quantification models into our analysis, such as DCE-Cross-Voxel Exchange, assessing their potential in producing refined, contextually-relevant biomarkers to further improve stratification4. References: 1. Patel, T.R. et al. A Comprehensive Review of MR Imaging Changes following Radiosurgery to 500 Brain Metastases. AM. J. Neuroradiol. 32, 1885–1892 (2011).2. Minniti, G. et al. Stereotactic radiosurgery for brain metastases: analysis of outcome and risk of brain radionecrosis. Radiat. Oncol. Lond. Engl.6, 48 (2011).3. Telera, S. et al. Radionecrosis induced by stereotactic radiosurgery of brain metastases: results of surgery and outcome of disease. J. Neurooncol.113, 313–325 (2013).4. Sinno, N. et al. Incorporating cross- voxel exchange for the analysis of dynamic contrast- enhanced imaging data: Pre-clinical results. Physics in Medicine & Biology67, 245013 (2022). Keywords: machine learning, radiomics, radiosurgery Digital Poster 4455 The Radiomic Fingerprint of the Marrow: Predicting Blood Toxicity before Lung Cancer Radiotherapy José Luis López Guerra 1 , Guillermo Canterla Casas 2 , María del Carmen Serrano Gotarredona 2 , Manuel Borrego Reina 1 , Begoña Acha Piñero 2 1 Oncología Radioterápica, Hospital Universitario Virgen del Rocío, Seville, Spain. 2 Teoría de la señal y comunicación, Universidad de Sevilla, Seville, Spain
Material/Methods: Imaging datasets were obtained from patients with confirmed TP/RN histopathological diagnoses prior to surgery, as part of an ongoing trial (NCT04244019). Retrospective cases from equivalent prior internal trials were included alongside prospective data to improve model robustness. The resulting dataset (n=55) incorporates diffusion, perfusion and permeability MRI (DWI-ADC, DSC-MRI and DCE-MRI) alongside standard anatomical MRI acquisitions (T1- weighted pre-/post-contrast, T2-weighted post- contrast) in BM arising from various primary cancers across a wide range of lesion volumes (<100 to >20,000 mm3). All acquired sequences were re- sampled to a common voxel space to promote feature stability. BM contours were interpolated to divide the clinical gross tumour volume into non-overlapping sub-compartments: the lesion core, edge, and periphery. Radiomic features were extracted per- volume and per-sequence, subsequently being used to train unique Random Forest classifiers for each sequence/volume grouping. Model performance was evaluated using the leave-one-out cross-validation method. ROC-AUC, sensitivity, and specificity scores were reported to compare model classification performance. Results: Our best-performing model, trained on features extracted from perfusion-weighted MRI within the lesion periphery (relative cerebral blood flow; DSC- MRI), achieved a cross-validated ROC-AUC of 0.829 (95% CI: 0.719–0.921; sensitivity: 0.760; specificity: 0.750; see Figure 1 below); the corresponding negative controls employing solely shape-based features achieved a ROC-AUC of 0.604 (95% CI: 0.470–0.733; sensitivity: 0.733; specificity: 0.630). The historical reference model (T1-weighted post-contrast) achieved a ROC-AUC of 0.596 (95% CI: 0.437–0.700; sensitivity: 0.700; specificity: 0.481).
Purpose/Objective: This study aims to predict the risk of acute
haematological toxicity (HT) before radiotherapy (RT) in lung-cancer patients by performing a multivariable analysis that integrates radiomics features extracted from segmented bone marrow (BM) structures in pre- RT computed tomography (CT) scans. Material/Methods: We analysed 698 lung cancer patients who underwent baseline thoracic CT scans and were treated with RT. BM masks were generated using a two-stage automated pipeline that had previously been validated. Then, 12,218 radiomics features were extracted from both the bone and the bone marrow, using both individual skeletal masks and global masks (whole skeleton and whole marrow). T-test and correlation filtering were then applied to reduce
Made with FlippingBook - Share PDF online