ESTRO 2021 Abstract Book

S1547

ESTRO 2021

Compostela, Spain; 3 Complexo Hospitalario Universitario de Santiago de Compostela, Department of Medical Physics, Santiago de Compostela, Spain; 4 Complexo Hospitalario Universitario de Santiago de Compostela, Department of Radiotherapy, Santiago de Compostela, Spain Purpose or Objective The increasing use of hypofractionation has motivated an intense debate on the validity of the linear-quadratic model (LQ) for large dose fractions. Variations with respect to small doses might arise either due to the sometimes called indirect cell death mechanisms, associated to vascular damage and immune-response activation, or due to a loss of effectiveness/damage saturation. In this work we investigate the possible contribution of indirect damage and damage saturation in response to SBRT/SRS treatments for early-stage NSCLC and brain metastases (BM). Materials and Methods We have constructed a dataset of 61 early-stage NSCLC and 51 BM dose-response cohorts of different fractionation schedules. Kaplan-Meier local tumor control probability (TCP) data are fitted to different models using a logistic function parametrized in terms of EQD2:

In order to produce realistic dose response curves, normalized gradients are fixed to match the steepness of clinical NSCLC and BM response curves (γ50=0.83 and 0.7, respectively) and the TCP is set to an upper limit of 0.95. EQD2model is calculated using the LQ, the linear-quadratic-linear (LQL), and phenomenological modifications of the LQ model to account for indirect cell damage (LQmod), with dose-dependent α-β terms of the following form:

To explore a variety of soft to strong dose dependencies, a’ and b’ are given values of 1/2, 1/3, 1 and 2. We compare the performance of the different models using the Akaike-Information-Criterion formalism (AIC) and log-likelihood ratio tests to check the significance of the results. We also study the stability of the results with changes in fitting parameters (α/β, accelerated proliferation kick-off times, and γ50) and perturbations on dose/TCP values. Results In NSCLC, using a typical α/β ratio of 10 Gy, a modified LQ-model involving a beta-term increasing with the square root of the dose (a=0, b>0 and b’=1/2), provides the best fit: AIC(LQmod)=344.6 vs AIC(LQ)=350.7, with p=0.004. Only the inclusion of very fast proliferation or low α/β values eliminates this superiority. For α/β~5Gy the behavior is inverted, and models with decreasing radiosensitivity with increasing dose, such as the LQL, become superior to the LQ. In this scenario, AIC(LQL)=332.8 vs AIC(LQ)=339.7, with p=0.003. In BM, the LQL model yields the best-fits, showing a decreased effectivity at large radiation doses: AIC(LQL)=701.6 vs AIC(LQ)=825.6, p<10-11. The outperformance of the LQL is observed for all fitting parameter values and dose/TCP perturbations explored. Conclusion The results for NSCLC are strongly dependent on the α/β value and may require further investigation, while those for BM seem to be clearly significant. Our observations may assist with the design of optimal radiotherapy treatments to avoid over- or under-treatment of these treatment sites. Purpose or Objective To implement a non-invasive workflow based on radiomics for the prediction of the grade of cervix cancer (CC) using DWI. Materials and Methods 39 patients with CC were collected from the medical records of our institution. For each patient, pretreatment DWI (with 3 b-values 0, 100, and 1000), manual delineation of the GTV and grade classification (17 with grade 2, 22 with grade 3) were available. Apparent Diffusion Coefficient maps ADC 100 and ADC 1000 corresponding to b-values 100 and 1000 respectively were generated using the mono-exponential model .The dataset was randomized and divided into training set (30 patients) and validation set (9 patients). Radiomic features were extracted from DWI (b-value = 0, 100, 1000) and ADC maps (ADC 100 and ADC 1000 ). 42 radiomic features were extracted from GTV’s volume of interest for each image volume. They were composed of 3 shape, 8 first order, and 31 texture features. Shape features describe the 3D geometric properties of the tumor. The first order features describe tumor intensity distribution. Texture features, which describe the intra-tumor heterogeneity, were extracted from the gray level co-occurrence (6), neighborhood grey-level different (3), grey-level run length (11), and grey-level zone length (11) matrices. Machine learning models (MLM) were composed of three blocs: preprocessing, selection and classification. Features preprocessing methods were TR_GAUSS (mapping data to have normal distribution), MINMAX_SC (linear scaling to [0, 1]), or ROBUST_SC (scaling to the interquartile range), STD_SC (removing the mean and scaling to unit variance), TR_UNIF (mapping to uniform distribution), L2_NORM (scaling samples to have unit norm). The selection methods of the most representative features were based on univariate statistical tests (Anova-F score (AF) or mutual information (MI)), or Support Vector Machine (SVM) model which performs class separation by hyperplanes. Binary classification, into grade 2 or 3, was by sparse linear model with elastic net regularization. MLM was assessed by the computation of the ROC curve, the area under the ROC curve (AUC) (measure of separability between classes for any threshold) and the accuracy (the fraction of correct predictions) for each grade. Results Table 1 shows AUC and accuracy on the training (AUC t , ACC t ) and validation (AUC v , ACC v ) sets obtained for each MLM. MLM with features processed by TR_GAUSS and selected with SVM gave the best results; (AUC t = 1, ACC t =1) and (AUC v = 0.81, PO-1818 Non-invasive grading of cervix cancer using diffusion weighted imaging S. aouadi 1 , S. Chandramouli 1 , T. Torfeh 1 , R. Hammoud 1 , N. Al-hammadi 1 1 Hamad Medical Corporation, National Centre for Cancer Care and Research, Doha, Qatar