ESTRO 2021 Abstract Book

S72

ESTRO 2021

Table1 shows the area under the curve of all the collected endpoints on the independent validation set in comparison with the clinical parameter based models. The clinical parameter based models were developed using a multi-layer perceptron. The best performed image based model was for regional recurrence prediction with an area under the curve (AUC) of 0.81, followed by the models for distant metastasis (AUC: 0.78) and local recurrence (AUC: 0.71).

Conclusion We developed and tested deep learning models based on pretreatment CT-scans for prognostic outcome prediction of head and neck cancer patients after curative (chemo)radiation. The CT image based deep learning models helped to improve the prediction performance on distant metastasis, local recurrence and regional recurrence. PH-0104 [18F]-FDG PET radiomics to predict survival in locally advanced cervical cancer and anal carcinomas S. Niyoteka 1 1 Gustave Roussy Cancer Campus, Departement of Radiation Oncology, Villejuif, France Purpose or Objective Locally advanced cervical cancer (LACC) and anal squamous cell carcinoma (ASCC) can both be caused by oncogenic human papillomaviruses (HPV), sharing many similarities in cancer biology. The purpose was to develop a regression model based on imaging features extracted from pre-treatment fluorine-18- fluorodeoxyglucose positron emission tomography ([18F]-FDG PET) images predicting survival, generalizable to In this study, cohorts from three different institutions were used. Two cohorts were constituted of LACC patients: 118 patients from Gustave Roussy Campus Cancer (GRCC, France), and 93 patients from St James’s University Hospital (SJUH, United Kingdom). The third cohort, including 90 patients from Institut du Cancer de Montpellier (ICM, France), corresponded to a set of ASCC patients. Patients were included based on the following criteria: i) histologically-confirmed LACC or p16+ASCC, ii) available pre-treatment [18F]-FDG PET scans. LACC and ASCC patients were treated by concurrent chemoradiotherapy +/- brachytherapy. First and second-order radiomics features (107) were extracted with Pyradiomics python package (version 3.0.1). To mitigate variability due to different scanners, we applied the ComBat technique for reducing batch effect to the extracted features. Four machine-learning (ML) regression models were built to predict overall survival (OS), progression-free survival (PFS), pelvic-free relapse survival (PFRS) and extra pelvic-free relapse survival (EPFRS) using harmonized imaging features as inputs (Figure 1). Each model was optimized in GRCC & ICM cohorts by excluding redundant features and by fine-tuning hyperparameters through ten-fold cross-validation repeated ten times. Each hyper-tuned model was tested on SJUH cohort. C-index of radiomics features-based ML models were compared to the performance of survival prediction models built with established prognostic factors in the literature: tumor SUV max , Metabolic Tumor Volume, Total Lesion Glycolysis, age, and lymph node status. HPV-induced cancers. Materials and Methods