ESTRO 2025 - Abstract Book

S3740

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Nantes, France. 14 Radiation Oncology, Institut de Cancérologie de Lorraine, Vandoeuvre les Nancy, France. 15 Radiation Oncology, Institut Cancer de Montpellier, Montpellier, France. 16 Research, LaTIM, Inserm, UMR 1101, Univ Brest, Brest, France. 17 Pathological Anatomy and Neuropathology, Aix-Marseille University, La Timone Hospital, Marseille, France. 18 Research, Institut de Neurophysiopathologie, CNRS - UMR 7051, Marseille, France. 19 Neuropathology, GHU Paris-Psychiatrie Et Neurosciences, Sainte-Anne Hospital, Paris, France. 20 Pathology, Toulouse University Hospital, Toulouse, France. 21 Research, Cancer Research Center of Toulouse (CRCT), INSERM U1037, Toulouse, France. 22 Research, Université Paul Sabatier, Toulouse, France. 23 Clinical Research and Innovation, Centre Léon Bérard, Lyon, France. 24 Pediatric Onco-Hematology, Institut d’Hématologie et d’Oncologie Pédiatrique, Lyon, France Purpose/Objective: Ependymoma (EPN) is a common childhood malignant brain tumor. Despite surgery and radiotherapy (RT), the 5 year relapse rate is 40% [1, 2]. Radiomics analysis has been applied for risk of relapse stratification using post contrast T1-WI [3]. However, EPNs have distinct features on diffusion-weighted imaging (DWI) with low apparent diffusion coefficient (ADC) values being related to tumor aggressiveness [4]. This study aimed to determine if DWI based radiomics could improve relapse-free survival prediction. Material/Methods: A total of 336 patients aged ≤22 with pathologically confirmed intracranial EPN between 2000 and 2021 and treated with postoperative RT were included from 13 clinical centers (NCT05151718). DWI data at diagnosis was available for 152 patients. Post contrast T1-WI was used to contour the entire tumor. Each ADC map was co-registered to the T1-WI and was preprocessed as described elsewhere [3]. Then, 109 Image Biomarker Standardization Initiative (IBSI)-compliant radiomics features were extracted using the LIFEx 7.3.0 package [5]. We also included clinical variables and visually derived characteristics [6]. The study's primary outcome was relapse-free survival (RFS) [3]. A radiomic signature (Rad-score) was built with radiomics features associated with RFS using an Elastic net regression [3]. Several models were developed including an integrative model incorporating Rad-score, clinical, and visual features. The performance of models was assessed using the concordance index (C-index). Results: After a median follow-up of 84.4 months, relapse occurred in 42.8% of patients. Univariate analysis showed that low RT dose (p=0.035), young age at RT start (p=0.008), and high number of resections before RT (p=0.024) were associated with worse RFS. We found an association between the Rad-score (built using 21 selected features) and RFS (hazard ratio [HR] = 1.17 [1.12; 1.23], p<0.001). For multivariate analysis, delay between diagnosis and RT, age at start of RT, RT dose, radiation technique and extent of resection were selected for clinical model and necrosis for visual model. Only young age (HR=0.92 [0.86; 0.99], p=0.035) was associated with worse RFS. In the integrative model, only the Rad-score was associated with RFS (HR=1.18 [1.13; 1.24], p<0.001). Finally, the integrative model seems better to evaluate RFS (C-index 0.75) than either the Rad-score (C-index 0.73) or the clinical (C-index 0.63) and visual (C-index 0.56) models. Conclusion: This study showed that combining DWI derived-radiomic, clinical, and visual features best predict RFS in EPN patients. Future work will focus on integrating molecular data, improving image harmonization with deep learning, and validating models externally.

Keywords: Relapse free survival, Ependymoma, Radiomics

References: 1- Ducassou A., et al. Int J Radiat Oncol Biol Phys. 2018 Sep 1;102(1):166-173. doi: 10.1016/j.ijrobp.2018.05.036. Epub 2018 May 24