ESTRO 2020 Abstract book

S888 ESTRO 2020

unsupervised, and features reduction methods were considered as shown in Figure 1. Lymph nodes involvement and ENE were extracted as a binary outcome. The radiomics-models were benchmarked to clinical-factors models including: age, gender, Tstage, LVI (Lymph vascular invasion), PNI (Perineural Invasion), and tumor thickness. Performances were evaluated using the accuracy and F1 scores, suggested in case of unbalanced datasets

associated dwell times were identical. When the 2 fx were added together and TL were generated, the probability of cure was found to be virtually 100% for all probable disease sites and all disease volume between 0.1cm 3 and 10cm 3 . When the exercise was repeated for the re-planed 1 fx of 19Gy, the probability of cure dropped dramatically from 78.6% for 0.1cm 3 to 30.4% for 10cm 3 disease volume. It took increasing prescription dose to 24Gy to restore the probability of cure profile similar to 27Gy given in 2 fx.

Results Radiomics features from T2 were in general more informative than features from T1 for predicting positive lymph nodes. In the validation cohort, the best T2- radiomics models had an accuracy of 0.70 [95% CI: 0.67, 0.71] and F1 score of 0.77 [0.75, 0.79]. The best clinical model combining clinical and pathological information about the primary tumour had an accuracy of 0.64 [0.63, 0.68] and F1 score of 0.74 [0.73, 0.76]. The combined model (radiomics+clinical) had an accuracy of 0.72 [0.71, 0.74] and F1 score of 0.76 [0.74, 0.78]. For predicting ENE, features from T1 were more informative. In the validation, the best T1-radiomics model had an accuracy of 0.61 [95% CI, 0.59,0.63] and F1 score of 0.70 [0.67,0.72]; the best clinical similar performances. The best combined radiomics + clinical model had an accuracy of 0.63 [0.61, 0.65] and F1 score of 0.72 [0.69, 0.74]. Conclusion This proof of concept study shows that quantitative imaging features combined with clinical factors for neck evaluation in oral cavity cancer patients may improve predictive accuracy. Further optimization and validation of extracted radiomic parameters are needed to extend the generalizability of the model. PO-1550 Radiomics characteristics correlate with immune activation and HPV status in head and neck cancer J.H. Oh 1 , E. Katsoulakis 2 , N. Riaz 3 , Y. Yu 3 , A. Apte 4 , J. Leeman 5 , N. Katabi 6 , L. Morris 7 , T. Chan 3 , V. Hatzoglou 8 , N. Lee 3 , J. Deasy 4 1 Memorial Sloan-Kettering Cancer Center, Medical Physics, New York- NY, USA ; 2 Veterans Affairs, Department of Radiation Oncology, Tampa, USA ; 3 Memorial Sloan Kettering Cancer Center, Department of Radiation Oncology, New York, USA ; 4 Memorial Sloan Kettering Cancer Center, Department of Medical Physics, New York, USA ; 5 Dana Farber Cancer Institute/Brigham and Women's Hospital, Department of Radiation Oncology, Boston, USA ; 6 Memorial Sloan Kettering Cancer Center, Department of Pathology, New York, USA ; 7 Memorial Sloan Kettering Cancer Center, Department of Surgery, New York, USA ; 8 Memorial Sloan Kettering Cancer Center, Department of Radiology, New York, USA Purpose or Objective We investigated whether radiomic characteristics on contrast enhanced computed tomography (CT) can predict biological differences in head-and-neck squamous cell cancer (HNSCC) patients. Material and Methods

Conclusion We have produced a proof of concept explanation on why, despite apparent BED equivalence of the prescription doses, the clinical outcome from 2 x 13.5Gy is significantly better than 1 x 19.5Gy. We believe that the TL model can produce insight and guidance into choosing a peripheral dose, not based on BED equivalence of the prescription dose but based on the equivalent performance at the level of the tumorlets, which can be customized based on the size of the index lesion, grade and the particular patient anatomy. PO-1549 Non-invasive prediction of lymph node risk in oral cavity cancer patients A. Traverso 1 , A. Hosni Abdalaty 2 , M. Hasan 2 , T. Tadic 3 , T. Patel 3 , M. Giuliani 2 , J. Kim 2 , J. Ringash 2 , J. Cho 2 , S. Bratman 2 , A. Bayley 2 , J. Waldron 2 , B. O'Sullivan 2 , J. Irish 2 , D. Chepeha 2 , J. De Almeida 2 , D. Goldstein 2 , D. Jaffray 3 , L. Wee 1 , A. Dekker 1 , A. Hope 2 1 Maastricht Radiation Oncology MAASTRO clinic, Radiotherapy, Maastricht, The Netherlands ; 2 Princess Margaret Cancer Centre, Department of radiation oncology, Toronto, Canada ; 3 Princess Margaret Cancer Centre, Radiation Medicine Programme, Toronto, Canada Purpose or Objective In oral cavity (OC) squamous cell cancer, the incidence of occult nodal metastases varies from 20% to 50% depending and tumour size and thickness. Besides clinical and histopathological factors, image-derived biomarkers may help estimate the probability of LN (lymph nodes) metastasis using a non-invasive approach to further stratify patients' need for neck dissection. We investigated the role of MR-based radiomics in predicting positive lymph nodes and ENE (extra nodal extension) in OC patients, prior to surgery Material and Methods 243 patients treated with curative intent neck dissection between 2003 and 2017 were considered in this study. 107 radiomics features were extracted from the manually delineated GTV for both T1/T2 sequences using PyRadiomics v2.1.2. Image resampling and normalization to further reduce differences in acquisition settings was performed prior to features xtraction. The dataset was split in 90% development (of which 70% training, 30% testing. Randomly bootstrapped 100 times) and 10% external validation for models' evaluation left apart. Features highly correlated with GTV (|Spearman| > 0.8) were removed prior to model development to avoid volume confounding effects. For model development, different machine learning algorithms, supervised and