S2416
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
treatment prediction of recurrence could enable risk- adapted treatment strategies and personalized follow- up. This study aimed to develop and compare traditional machine learning (ML) models based on radiomics and clinical variables with deep learning (DL) approaches utilising foundation model–derived representations from pre-treatment MRI for recurrence prediction in ASCC. Previous radiomics studies on ASCC recurrence prediction have not included an independent external validation set, limiting generalizability [1,2]. In contrast, this analysis incorporated an external validation cohort to ensure robust evaluation. Material/Methods: Pre-treatment T2-weighted MRI scans from 184 patients with ASCC treated with definitive chemoradiotherapy between 2004 and 2023 were retrospectively analysed. Data were collected from five German Cancer Consortium (DKTK) centers and one external center (Innsbruck, Austria). Two segmentation definitions were evaluated: the Gross Tumour Volume (GTV) and an extended region including the Anal Canal (GTVAC). Radiomics features were extracted from both regions and harmonized across institutions to account for scanner and acquisition variability. Clinical variables included age, sex, HIV status, T-status, and N- status. Three recurrence-related endpoints were analyzed: local recurrence, locoregional recurrence, and salvage surgery. Data preprocessing included feature selection using Minimum Redundancy Maximum Relevance (mRMR), class balancing via SMOTE and undersampling, feature normalization and harmonization. Three traditional ML models—Extreme Gradient Boosting (XGB), Random Forest (RF), and Support Vector Machine (SVM)—were trained on clinical, radiomics, and combined data. Their performance was compared with a deep learning framework based on the CT-trained foundation model FMCIB [3], fine-tuned with a multilayer perceptron (MLP) head for MRI-based recurrence prediction. Results: Among the ML models, the RF classifier achieved the best performance in external validation using radiomics data from the GTVAC region, reaching an AUROC of 0.78, while the clinical model achieved an AUROC of 0.60, with the best performing model overall based on the AUROC being the FMCIB + MLP model using the GTV region, which achieved an AUROC of 0.79.
A., 2020. On explaining machine learning models by evolving crucial and compact features. Swarm and Evolutionary Computation, Volume 53, p. 100640. Keywords: Survival Prediction, bone metastases, XAI
Digital Poster 678 Multi-center MRI radiomics for recurrence prediction in anal squamous cell carcinoma using machine learning and a foundation model-based approach Lucas Zander 1,2 , Ahmed Mohamed 1,3 , Katharina Riegger 1 , Óscar Llorian Salvádor 3,1 , Lothar Richter 3 , Daniel Martin 4,5 , Claus Rödel 4,5 , Henning Schäfer 6,7 , Anca-Ligia Grosu 6,7 , Cihan Gani 8,9 , Maximilian Niyazi 8,9 , Goda Kalinauskaite 10,11 , Daniel Zips 10,11 , Elisa Thomas 12,13 , Mechthild Krause 12,13 , Samuel Vorbach 14 , Ute Ganswindt 14 , Burkhard Rost 3 , Hendrik Dapper 15,1 , Denise Bernhardt 1,2 , Stephanie E. Combs 1,2 , Jan C. Peeken 1,2 1 Department of Radiation Oncology, TUM University Hospital, Technical University Munich (TUM), Munich, Germany. 2 German Cancer Consortium (DKTK), Partner Site Munich, Munich, Germany. 3 Informatics 12, Chair for Bioinformatics, Technical University Munich (TUM), Munich, Germany. 4 Department of Radiotherapy and Oncology, University Hospital Frankfurt, Goethe University Frankfurt, Frankfurt, Germany. 5 German Cancer Consortium (DKTK), Partner Site Frankfurt, Frankfurt, Germany. 6 Department of Radiation Oncology, University Hospital of Freiburg, Albert Ludwig University of Freiburg, Freiburg, Germany. 7 German Cancer Consortium (DKTK), Partner Site Freiburg, Freiburg, Germany. 8 Department of Radiation Oncology, University Hospital and Medical Faculty, Eberhard Karls University Tübingen, Tübingen, Germany. 9 German Cancer Consortium (DKTK), Partner Site Tübingen, Tübingen, Germany. 10 Department of Radiation Oncology, Charité – Universitätsmedizin Berlin, Berlin, Germany. 11 German Cancer Consortium
(DKTK), Partner Site Berlin, Berlin, Germany. 12 Department of Radiotherapy and Radiation
Oncology, Medical Faculty Carl Gustav Carus, Technical University of Dresden, Dresden, Germany. 13 German Cancer Consortium (DKTK), Partner Site Dresden, Dresden, Germany. 14 Department of Radiation Oncology, Medical University of Innsbruck, Innsbruck, Austria. 15 Department of Radiation Oncology, Cyberknife and Radiotherapy, University Hospital Cologne, Cologne, Germany Purpose/Objective: Anal squamous cell carcinoma (ASCC) is primarily treated with definitive chemoradiotherapy, yet disease recurrence remains a clinical challenge. Reliable pre-
Table 1: Performance comparison (AUROC and F1@0.5 with 95% CI) Conclusion:
Made with FlippingBook - Share PDF online