S2494
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Results: The top-performing convolutional neural network, DenseNet169, achieved a mean AUC of 0.75 (range: 0.71-0.81) for distinguishing high versus low SIR in a 5- fold cross-validation (Figure 1). In the oropharyngeal subset used for multivariable analysis, 1,437 patients with 91 recurrences were analysed, with a median (censored) follow-up of 4.4 years. Probability estimates from the CNN-classified SIR remained an independent predictor of recurrence, yielding a hazard ratio (HR) (per 0.10 increase in probability estimate) of 1.67 (95%CI: 1.25-2.24, p<.005), in multivariate analysis (Figure 2). The SIR probability estimates met linearity and proportional hazards assumption. Univariate analyses showed strong prognostic for all endpoints: RFS had an HR of 9.57 (95%CI: 2.56-35.76, p<.001); OS had an HR of 3.94 (95%CI: 1.90-8.17, p<.001); and DFS had an HR of 2.53 (95%CI: 1.25-4.72, p=.009). Kaplan- Meier curves demonstrated clear separation with log- rank p <.001. The three-year RFS was 91.9%, OS was 77.3% versus 74.7%, and DFS was 70.3% versus 68.2%. Prior analyses in an independent randomized trial cohort showed concordant trends. Conclusion: A CT-derived immune phenotype score validated on the RADCURE dataset provides independent prognostic information for recurrence and survival in oropharyngeal cancer. These results support the prospective use of CT immune phenotyping for risk stratification and individualized treatment. References: 1. Welch, M.L., et al., RADCURE: An open-source head and neck cancer CT dataset for clinical radiation therapy insights. Med Phys, 2024. 51(4): p. 3101– 3109.2. Altunbulakli, C., et al., Targeted spatial proteomic analysis of CD8+ T- and myeloid cells in tonsillar cancer. Frontiers in Oncology, 2023. 13.3. Zuley, M.L., et al., The Cancer Genome Atlas Head- Neck Squamous Cell Carcinoma Collection (TCGA- HNSC). 2016: The Cancer Imaging Archive.4. (CPTAC), N.C.I.C.P.T.A.C., The Clinical Proteomic Tumor Analysis Consortium Head and Neck Squamous Cell Carcinoma Collection (CPTAC-HNSCC) (Version 18) 2018: The Cancer Imaging Archive. Keywords: Immune biomarker, deep learning, Oropharyngeal Digital Poster 4821 Low-dose distribution predicts radiation pneumonitis in esophageal cancer: impact of gastroesophageal junction involvement Camila Souza dos Santos 1 , Victor Gurgel da Fonseca 2 , Alexandre dos Santos Gomes 1 , Rachele Grazziotin Reisner 1
extracted relevant radiological features for surgical and SBRT patients, resulting in favorableclassification performance. Our model could be a promising tool to treat patients woPC with SBRT,avoiding potential risks associated to biopsy. Keywords: Fundation Model, Lung , SBRT
Digital Poster 4818
Deep learning CT-derived immune phenotypes predict recurrence and survival in the RADCURE oropharyngeal cancer cohort André Haraldsson 1 , David Askmyr 2 , Can Altunbulakli 3 , Mikael Nilsson 4 , Jean-philippe Pignol 1 , Christian Jamtheim Gustafsson 1 , Anton Linnér 1 , Malin Lindstedt 3 , Lennart Greiff 2 , Maria Gebre-Medhin 5 1 Radiation physics,Dept of hematology, oncology, radiation physics, Skåne university hospital, Lund, Sweden. 2 Clinical Sciences, Lund university, Lund, Sweden. 3 Immunetechnology, Lund uinversity, Lund, Sweden. 4 Centre for mathematical sciences, Lund university, Lund, Sweden. 5 Dept of hematology, oncology, radiation physics, Skåne university hospital, Lund, Sweden Purpose/Objective: Immune phenotypes based on the spatial distribution of CD8+ T cells in solid cancers carry important prognostic information but extracting the information is invasive. We trained and validated a deep learning model on head and neck patients to assess T cell infiltration based on planning CT and evaluated its association with survival in the RADCURE dataset[1]. Material/Methods: In the development phase, 104 patients with tonsillar cancer and paired immunohistochemistry[2] were used to derive an immune stroma-infiltrate-ratio (SIR) from spatial T cell density. We trained a DenseNet169 to assess infiltration from planning CTs. Additionally, the performance on bulk mRNA was validate on a subset of TCGA[3] and CPTAC[4] patients and survival prognostic on RADCURE patients who underwent curative-intent radiotherapy with or without chemotherapy. Survival outcomes were assessed using the model’s probability estimate, including recurrence-free survival (RFS), overall survival (OS), and disease-free survival (DFS), measured from treatment start to event or censoring. The primary analysis entered the SIR probability as a continuous covariate, into Cox models adjusted for HPV/p16 status, tumour stage, dose, and chemotherapy. Linearity and proportional hazards assumptions were evaluated. For survival stratification, patients were dichotomised by an F1-optimised cutoff.
Made with FlippingBook - Share PDF online