ESTRO 2020 Abstract book

S446 ESTRO 2020

(accuracy=95%), showing good results in terms of sensitivity (86%), specificity (97%), NPV (97%) and PPV (86%), as shown in Fig.1. The value of 23Gy represents the optimal BED level to perform prediction; ERI estimated at different BED levels showed lower AUC: 0.69 (0.41-0.93) for ERI 12Gy , 0.87 (0.73-0.98) for ERI 36Gy and 0.78 (0.56-0.96) for ERI 45Gy

LARC, using a third external independent extraeuropean validation cohort. Material and Methods The original multivariable model was based on 4 covariates: clinical T and N staging and 2 radiomics features extracted from vendor independent staging 1.5 T MRI (entropy of the image and skewness of the ROI grey level histogram filtered). The considered binary outcome was pCR achievement. The model was realized using a single center training set of 162 patients (pts) and 2 external validation sets for a total of 59 pts provided by other 2 European centers. All the pts received long course NAD CRT. The first validation of the model confirmed its stability on both internal and external validation, with AUC values of 0.73 and 0.75 respectively. In order to test model’s replicability, a third extraeuropean validation cohort has been enrolled using real world data of a Chinese institution including both 1.5 and 3 T staging MRI images, and pts receiving both long and short course NAD CRT. Validation was performed via Receiving Operator Characteristics (ROC) curve analysis and classification matrix metrics at the ROC Youden index cut-off point. ROC AUC and classifications matrix accuracy, specificity, sensitivity, negative predictive value, positive predictive value, and Kappa statistics were taken as validation metrics. Results 60 Chinese pts were enrolled in the extraeuropean validation cohort, showing a pCR occurrence of 16% (10 cases). 27 and 33 image datasets were collected from 1.5 T and 3 T scanners. 21.9% and 78.1% of the pts received short and long course NAD treatment respectively. Proper Laplacian of Gaussian sigma and pixel spacing values have been taken into account for model applicability. A ROC AUC of 0.83 was achieved on the whole Chinese validation dataset (see fig. 1). Classification matrix on the testing dataset best cut-off point according to Youden index: 0.85 accuracy, 0.90 specificity, 0.60 sensitivity, 0.92 negative predictive value, 0.55 positive predictive value, 0.48 Kappa statistics.

Conclusion This study confirmed ERI TCP as a potentially powerful biomarker able to early predict the pCR after radio- chemotherapy for LARC. This performance was confirmed with MR images obtained during MRgRT, acquired with different imaging sequences and magnetic field strength than those used in the original training cohort. These results, if confirmed in larger cohorts, could facilitate the clinical implementation of ERI TCP in personalized protocols for LARC PH-0716 Radiomics pCR predictive model in rectal cancer: an intercontinental validation on real world data L. Boldrini 1 , J. Lenkowicz 1 , L.C. Orlandini 2 , N. Dinapoli 1 , G. Yin 2 , D. Cusumano 1 , C. Casà 1 , Q. Peng 1 , G. Chiloiro 1 , M.A. Gambacorta 1 , J. LANG 2 , V. Valentini 1 1 Fondazione Policlinico Universitario "A. Gemelli" IRCCS, Diagnostica per Immagini- Radioterapia Oncologica ed Ematologia, Roma, Italy ; 2 Sichuan Cancer Hospital & Institute- Sichuan Cancer Center- School of Medicine- University of Electronic Science and Technology of China, Department of Radiation Oncology, Chengdu, China Purpose or Objective Pathological complete response (pCR) prediction in locally advanced rectal cancer (LARC) undergoing neoadjuvant (NAD) chemoradiotherapy (CRT) represents a promising field of investigation. Unfortunately, the predictive models published so far often lack of replicability and reliable external validations. Aim of this study was to evaluate the replicability of an already published multivariable pCR predictive model in

Conclusion Despite the introduction of significant different variables (i.e. ethnic origin, MRI field strength, type of NAD treatment) the proposed model appeared to be replicable and stable on a real world data extraeuropean patients cohort. The obtained promising results encourage to further investigate the application of radiomics modeling in the frame of multivariable decision support systems for LARC.