S2480
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
References: [1] Kaplan J, McCandlish S, Henighan T, Brown TB, Chess B, Child R, et al. Scaling Laws for Neural Language Models 2020. https://doi.org/https://doi.org/10.48550/arXiv.2001.08 361. Keywords: Deep Learning, Scaling Law, Survival Analysis Digital Poster 4015 Exploring radiomics-based integrated molecular- morphologic classification in high-grade meningioma treated with proton therapy: preliminary insights Giulia Fontana 1 , Sara Lillo 1,2 , Alessia Bazani 1 , Sithin Thulasi Seetha 1 , Luca D'Ambrosio 1 , Alberto Iannalfi 1 , Silvia Molinelli 1 , Ester Orlandi 1,3 1 Clinical, (CNAO) National Center for Oncological Hadrontherapy, Pavia, Italy. 2 Internal Medicine and Therapeutics, University of Pavia, Pavia, Italy. 3 Clinical, Surgical, Diagnostic, and Pediatric Sciences, University of Pavia, Pavia, Italy Purpose/Objective: To explore the prognostic effectiveness of the radiomics features associated with the novel integrated molecular-morphologic (iMM[1]) classification for high-grade meningioma treated with proton therapy. Material/Methods: Patients affected by WHO II or III meningioma, treated with curative proton therapy at our Institution between 2014 and 2021, were included in the study. Kertels et al. developed a radiomics signature from multicenter magnetic resonance imaging data (contrast-enhanced T1-weighted, T1ce), effectively stratifying high-grade meningioma into low-, medium-, and high-risk groups based on the iMM classification[2]. While their fitted model is not public, the reported radiomics features with the highest importance for iMM prediction were evaluated (6 shape, 2 first-order, and 2 texture) on our population. The pre-treatment T1ce sequences were retrieved from the institutional repository, along with the planning gross tumor volume. The selected features were extracted after z-score normalization. Univariable Cox proportional-hazard models were fitted to explore the impact of the selected radiomics features on the local disease progression. Finally, the area under the time-dependent Receiver Operating Characteristic curve (time-AUROC) was computed as the feature’s accuracy metric for local relapse prediction at three-years[3]. A significance level of 0.05 and a cut-off of 0.7 for time-AUROC were set.
Cindex(N)=aNb, where N is the dataset size, was used to fit the Cindex validation performances. Results: Figure 1A shows the mean 10-fold cross-validation Cindex per training epoch for the different dataset sizes. Figure 1B shows the mean per-fold-peak cross- validation Cindex at different training set sizes: 0.575 [IQR:0.563-0.594], 0.583 [IQR:0.578-0.591], 0.607 [IQR:0.593-0.611], 0.617 [IQR:0.604-0.6321], for 200, 500, 1000, and 2000 patients respectively. Also plotted in Figure 1B is the power law regression with estimated parameters: â = 4.83 × 10-1±6.70 × 10-2, b̂ = 3.24 × 10-2±2.12 × 10-2, with R2 = 0.962. Extrapolating from our scaling law, a cohort of 105 patients is projected to yield a model Ĉ index of approximately 0.7. Conclusion: Our results suggest a need to re-evaluate data requirements for AI in radiotherapy outcome prediction. The power scaling law relationship we describe estimates that reliable OS prediction models likely require datasets in the tens of thousands of patients to reach clinical utility, a threshold that exceeds the scope of most current efforts. Future work will focus on refining scaling law estimation, and
evaluating performance gains by including pathological, tumour biology, and concurrent anticancer therapy information.
Made with FlippingBook - Share PDF online