S838
Clinical - Lung
ESTRO 2026
intermuscular fat);calculate blood dose metrics reflecting circulating dose exposure (EDIC);compute coronary and aortic calcium scores;estimate the diameters of the aortic arch and descending aorta normalised by body surface;assess Low Attenuation Area (LAA) normalised to lung volumes, describing the percentage of emphysema.All extracted variables were integrated into a dedicated analysis pipeline. After feature selection using the Variance Inflation Factor, mutual information and permutation test, three machine learning models (LightGBM, XGBoost, and Random Forest) were trained and tested using a 70/30 data split. Model performance and feature importance were evaluated through AUC and SHAP analysis. To enhance interpretability, the multidimensional feature space was reduced using UMAP, and clustering was performed with HDBSCAN. Cluster-specific characteristics and survival rates were subsequently analysed. Results: OS2 was 56.7%. After preliminary filtering based on prior dosimetric predictors (EUD n=0.1 for the left anterior descending artery(LAD), mean dose to the inferior vena cava, and EDIC) [1,2], 23 variables, including clinical, treatment, and quantitative imaging features, were retained. Feature selection identified 6 key predictors of 2-year overall survival, ranked by SHAP importance (Figure1): EDIC, visceral fat volume, PTV, aortic Agatston calcium score, LAA and EUD(n=0.1) for the LAD.Among the evaluated models, Random Forest achieved the best performance, with AUC values of 0.83 (training), and 0.65 (test). UMAP– HDBSCAN clustering identified 8 distinct patient groups characterised by 2 dose metrics, 1 tumor variable and 3 image-derived factors.The 2-year death rates are reported in Figure 2.
Thymus dose and patients’ outcomes, and thereby their role as routinely implemented dose-guided OARs. Keywords: autosegmentation, side-effects, outcome- prediction Digital Poster Highlight 4295 Quantitative imaging and dosimetric signatures define survival clusters and inform dose optimisation in locally advanced NSCLC Alessandra Catalano 1 , Eliana Gioscio 1 , Francesco Dionisi 2 , Matteo D'andrea 3 , Lorenzo PLacidi 4 , Luca Boldrini 5 , Livia Marrazzo 6 , Emanuela Olmetto 7 , Patrizia Ciammella 8 , Andrea Botti 9 , Tiziana Rancati 1 , Anna Cavallo 10 , Francesco Pisani 1 , Silvia Meroni 10 , Albina Allajbej 11 , Claudia Sangalli 11 , Andrea Filippi 11 , Sara Broggi 10 , Roberta Tumminieri 11 , Alessandro Cicchetti 1 1 Data Science Unit, Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy. 2 Radiation Oncology, IRCCS Istituto Tumori Regina Elena, Rome, Italy. 3 Medical Physics, IRCCS Istituto Tumori Regina Elena, Rome, Italy. 4 Medical Physics, IRCCS Fondazione Policlinico Universitario Agostino Gemelli, Rome, Italy. 5 Radiation Oncology, IRCCS Fondazione Policlinico Universitario Agostino Gemelli, Rome, Italy. 6 Medical Physics, zienda Ospedaliera Universitaria Careggi, Florence, Italy. 7 Radiation Oncology, zienda Ospedaliera Universitaria Careggi, Florence, Italy. 8 Radiation Oncology, Azienda Unità Sanitaria Locale - IRCCS Reggio Emilia, Reggio Emilia, Italy. 9 Medical Physics, Azienda Unità Sanitaria Locale - IRCCS Reggio Emilia, Reggio Emilia, Italy. 10 Medical Physics, Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy. 11 Radiation Oncology, Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy Purpose/Objective: To evaluate dosimetric, cardiovascular, and metabolic risk factors influencing 2-year overall survival (OS2) in patients with locally advanced non-small cell lung cancer treated with curative-intent radiotherapy. Material/Methods: This multicentric retrospective study included 388 patients treated across six institutions between 2014 and 2022. In addition to radiotherapy, 30% received chemo-immunotherapy, 52% received chemotherapy, 5% received immunotherapy, and 13% received radical RT alone. Among patients receiving chemotherapy, 33% were treated sequentially, and 63% were treated concomitantly. Clinical, treatment, and longitudinal survival data were collected.Patient DICOM were processed to:automatically contour organs of interest, including cardiac substructures, and extract corresponding dose-volume parameters;derive metabolic volumes (visceral fat, skeletal muscle, and
Made with FlippingBook - Share PDF online