ESTRO 2025 - Abstract Book

S3819

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

3352

Digital Poster Exploring the performance gap between GTV and radiological lesion-based radiomics predicting adenoid cystic carcinoma progression after proton therapy Giulia Fontana 1 , Giulio Di Ciaccia 1 , Sithin Thulasi Seetha 1,2 , Cristina Fichera 1,2 , Maria Bonora 1 , Luca Trombetta 1 , Lorena Levante 1,2 , Barbara Vischioni 1 , Silvia Molinelli 1 , Sara Imparato 1 , Ester Orlandi 2,1 1 Clinical Department, (CNAO) National Center for Oncological Hadrontherapy, Pavia, Italy. 2 Department of Clinical, Surgical, Diagnostic and Pediatric Sciences, University of Pavia, Pavia, Italy Purpose/Objective: Local and distant relapse prediction in adenoid cystic carcinoma (ACC) after proton therapy (PT) may guide optimized treatment planning, hence leading to better outcomes. Computed Tomography (CT) radiomics models showed their value as predictive tool in radiotherapy, however lesion segmentations is a crucial step being time consuming and difficult to standardize. Our study aimed to explore the performance difference of CT-based radiomics predicting ACC progression after PT, relying on planning GTV or pre-treatment dedicated lesion contouring (RADT). Material/Methods: Planning-CT, RTStruct and follow-up radiological data of ACC reporting a macroscopic disease before PT were retrieved. Hence, GTV were polished to include only voxels within the soft tissue range (Hounsfield Unit (HU): [-200; +600]), as well a radiologist contoured the RADT on the planning-CT. 107 original radiomics features were extracted with Pyradiomics package, and afterwards globally standardized. A two-step feature selection process was implemented, firstly removing low-variance (10 -3 ) and highly correlated variables (Spearman Rho≥0.85), hence using a backward step-wise approach optimized for the selected classifiers based on area under the Receiver Operating Characteristic curve (AUROC). Logistic regression (LR), linear support vector machines (l-SVM), extreme gradient boosting (XGB), and random forest (RF) models were trained with two to five selected features. The mean AUROC 5 fold cross-validation was used to evaluate models’ performance, and DeLong’s test was used to explore statistically significant differences between the best GTV- and RADT-based radiomics models (α=0.05). Confidence intervals (CI) and bias-corrected p-value for AUROC were computed with bootstrap (n=1000). Results: 56 patients (41% males, and a median age of 61.2 years-old) affected by ACC were included in our study. 9 and 18 patients experienced a local and/or distant relapses, overall accounting for 20 patients experiencing at least one progression event. The GTV and RADT filtered features were 20 and 25 after the first-step of the feature selection process, while respectively logistic regression and XGB reported the highest performance, with 5 input features. The achieved AUROC were 0.827 (bootstrap CI: 0.69 – 0.98, p<0.0001) and 0.839 (0.61 – 0.92, p<0.0001) for GTV and RADT, respectively, with no statistically significant differences (p=0.949). Conclusion: CT-based radiomics models predicting ACC progression after PT reported comparable performances with respect to the considered lesion segmentation and despite the number of input features. In the analyzed radiomics setting, GTV showed the potential to represent a valuable, easier to standardize and less time-consuming lesion segmentation compared to radiological contouring. Comprehensive studies with larger datasets are warranted to confirm our findings.

Keywords: segmentation, Computed Tomography, proton therapy