ESTRO 2021 Abstract Book

S504

ESTRO 2021

head and neck squamous cell carcinoma (HNSCC) undergoing primary radiochemotherapy (pRCTx); and (ii) if the expression of gene signatures related to different biological processes associated with tumour radioresistance can be represented by CT-derived radiomic features Materials and Methods A multicentre retrospective cohort of 206 patients with locally advanced HNSCC was divided into training and validation sets (123 and 83 patients, respectively, Figure a). Gene expression profiling of tumour biopsies was performed using GeneChip ® human transcriptome arrays (Affymetrix). Radiomics features were derived from the primary tumour volume (GTV) on the treatment-planning CT scan. (i) Differentially-expressed genes were selected for 2-year LRC using a significant fold change ≥1.2 followed by clustering and elastic-net feature selection. Three Cox models were developed and validated for LRC prognosis based on radiomics features alone, gene expressions alone, and a combination of both datasets. (ii) Six gene signatures related to DNA repair, hypoxia, radiosensitivity, immune response and epithelial mesenchymal transition were selected from the literature and used to classify patients into two risk groups. Logistic regression was applied to predict the patient classification according to each signature using radiomic features. The prognosis of LRC was evaluated by the concordance-index (C-Index) with 95 % confidence interval (CI) and by the p-value from the log-rank test for stratified patient groups. Classifiers were assessed based on the area under the curve (AUC) and its (i) Combining radiomic features and gene expressions revealed a higher validation performance (C-Index=0.67 [0.55-0.75]) than radiomic features alone (0.61 [0.51-0.71]) or gene expressions alone (0.61 [0.51-0.70]). The final selected signature contained three radiomic features (including GTV) and two genes (NFIA, GPRC5D). In patient stratification (Figure b,c), a significant p-value was observed for the combined signature (p=0.002). (ii) Overall, radiomic features showed low correlations to the selected gene classifiers. Only for one signature related to epithelial-mesenchymal transition [2] ,a moderate AUC was observed in validation (0.60 [0.50- 0.71], Figure d). Conclusion Combining radiomic features with gene expression data showed an improved prognostic value for LRC in patients with locally advanced HNSCC after pRCTx. CT-based radiomic features and gene classifiers were weakly correlated, i.e., presented with independent information. Further validation of the biomarkers is planned. References [1] Rabasco Meneghetti A., et al. Clin Transl Radiat Oncol (2021) 26:62–70. [2] Chung CH, et al. Cancer Res 2006:8210-8. 95% CI. Results

OC-0639 Towards personalised treatment for prostate cancer: biology improves image-based data mining models J. Shortall 1 , E. Vasquez Osorio 1 , A. McWilliam 1 , A. Green 1 , A. Choudhury 2 , P. Hoskin 3 , C. West 1 , B. Lane 1 , T. Elumalai 3 , N. Thiruthaneeswaran 1 , B. Bibby 1 , R. Rodrigues Pereira 4 , M. van Herk 1 1 The University of Manchester, Faculty of Biology, Medicine and Health, Division of Cancer Sciences,