S198
Clinical - Biomarkers of clinical response
ESTRO 2026
local progression with an AUC of 0.667, while the radiomic model achieved an AUC of 0.791. The combined clinical–radiomic “multi-expert” model reached an AUC of 0.861, with an accuracy of 0.846 and an F1-score of 0.830, demonstrating superior predictive performance. These findings highlight the benefit of integrating radiomic and clinical features to identify patients at higher risk of local progression after chemoradiotherapy. Conclusion: Using a Mixture of Experts model to integrate radiomic and clinical features substantially enhances the predictive accuracy of local progression in LAPC. This hybrid approach demonstrates the feasibility and clinical potential of using AI-based radiomics to refine prognostic assessments and inform personalised treatment strategies for pancreatic cancer patients. However, further external validation is required to confirm its robustness and generalisability across independent cohorts. Keywords: AI, PDAC, radiomic MRI-based radiomics for predicting pathological response to preoperative radiotherapy in myxoid liposarcoma Silvia Ruggeri 1 , Mauro Loi 1 , Daniela Greto 1 , Sebastiano Paolucci 2 , Giuliana Roselli 3 , Roberto Scanferla 4 , Annarita Palomba 5 , Francesco Muratori 4 , Guido Scoccianti 4 , Filippo Nozzoli 5 , Marco Banini 1 , Olga Ruggieri 1 , Niccolò Bertini 1 , Elisabetta Neri 4 , Serena Puccini 4 , Cinzia Tramonti 2 , Monica Mangoni 1 , Linda Calistri 3 , Vittorio Miele 3 , Domenico Andrea Campanacci 4 , Lorenzo Livi 1 1 Radiation Oncology Department - Careggi Hospital, University of Florence, Florence, Italy. 2 Department of health Physics, University of Florence, Florence, Italy. 3 Department of Radiology - Careggi Hospital, Digital Poster 4692 University of Florence, Florence, Italy. 4 Department of orthopaedic Oncology and Reconstructive Surgery - Careggi Hospital, University of Florence, Florence, Italy. 5 Unit of Histopathology and Molecular Diagnostic - Careggi Hospital, University of Florence, Florence, Italy Purpose/Objective: Myxoid liposarcoma (MLPS) is a radiosensitive soft tissue sarcoma, yet the degree of response to preoperative radiotherapy (RT) varies considerably. Identifying non-invasive imaging biomarkers capable of predicting histopathological response before treatment could improve patient stratification and support personalised therapeutic strategies. This study investigated whether MRI-based radiomic features can predict pathological response to neoadjuvant RT in MLPS.
Digital Poster 4359 AI-Driven Radiomic Modeling Improves Prediction of Local Progression in Pancreatic Ductal Adenocarcinoma Treated with Chemoradiotherapy Gabriele D'Ercole 1 , Michele Fiore 1,2 , Alice Tretola 2 , Ermanno Cordelli 3 , Gian Marco Petrianni 1 , Pasquale Trecca 1 , Edy Ippolito 1,2 , Paolo Soda 3,4 , Sara Ramella 1,2 1 Operative Research Unit of Radiation Oncology, Fondazione Policlinico Campus Bio-Medico di Roma, Roma, Italy. 2 Research Unit of Radiation Oncology, Università Campus Bio-Medico di Roma, Roma, Italy. 3 Unit of Computer Systems and Bioinformatics, University Campus Bio-Medico of Rome, Roma, Italy. 4 Department of Diagnostics and Intervention, Radiation Physics, Biomedical Engineering, Umeå University, Umeå, Sweden Purpose/Objective: Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest types of cancer, with limited operability and poor survival rates. Identifying early predictors of disease progression is crucial for personalising treatment strategies. The aim of this study was to develop a clinical–radiomic model that integrates imaging and clinical data using a Mixture of Experts (MoE) approach, with the intention of predicting local progression (LP) and improving outcome stratification in patients with borderline resectable or locally advanced pancreatic cancer (LAPC) who are undergoing chemoradiotherapy (CRT). Material/Methods: A total of 86 consecutive patients with LAPC or borderline resectable PDAC, treated with CRT with or without induction chemotherapy, were retrospectively analyzed. Contrast-enhanced CT images in the venous phase were segmented to include the gross tumor volume. From these images, 1,595 radiomic features were extracted using PyRadiomics. Clinical and radiomic datasets underwent preprocessing (normalization, imputation, one-hot encoding). Two predictive models were developed: a Random Forest classifier for clinical features and a K-Nearest Neighbors classifier for radiomic features. These were subsequently integrated using a Mixture of Experts (MoE) ensemble to leverage complementary information. ROC analysis and F1-score were computed to evaluate predictive performance for LP. Results: The median age of the population was 64 years (range 36–79), with 85% of tumors located in the pancreatic head. The median overall survival (OS) for all patients was 14.2 months, and the median progression-free survival (PFS) was 9.5 months. Patients who underwent surgical resection showed significantly longer OS compared to those who did not (54.5 vs. 11.8 months, p < 0.01). The clinical model predicted
Made with FlippingBook - Share PDF online