ESTRO 38 Abstract book

S264 ESTRO 38

DFS (91% vs 65%, p=0.03) and overall survival (96% vs 77%, p=0.06). Conclusion MRI assisted external radiation and brachytherapy results in excellent 7-year disease free and overall survivals in patients with post surgical vaginal recurrences OC-0509 MRI radiomics analysis for predicting prognosis of cervical cancer after definitive radiotherapy A. Takada 1 , H. Yokota 1 , M. Watanabe 1 , T. Horikoshi 1 , T. Uno 1 1 Chiba University, Department of Radiology, Chiba, Japan Purpose or Objective Although many clinical prognostic factors for uterine cervical cancer have been reported, prediction accuracy is still insufficient. Radiomics is a method to construct predictive machine learning models for treatment prognosis or benign/malignant using many image features, which are extracted from a lesion site on medical images such as MRI, CT, and PET. However, machine learning is theoretically weak for differences in imaging acquisition conditions. The purpose of this study was to investigate whether radiomics analysis with MRI images of different scanners can predict the prognosis in cervical cancer after Totally 107 patients (58.8 ± 13.5 years) were included. All underwent MRI of 16 multi-center scanners and definitive radiotherapy for cervical cancer between April 2012 and March 2016. Spatial resolution of MRI was converted into 0.5 x 0.5 x 5.0 mm for T2-weighted image (T2WI), and 1.0 x 1.0 x 5.0 mm for diffusion-weighted image (DWI) and apparent diffusion coefficient map (ADC). A radiologist delineated the volume of interest (VOI) within each tumor region on axial T2WI, DWI and ADC map (VOI1). In addition, VOI2 was created by expanding the entire circumference by 4 mm. Using open-source software (LIFEx: https://www.lifexsoft.org), two intensity rescaling methods were applied; i.e. relative rescaling for T2WI and DWI, and relative and absolute rescaling for ADC. The Relative method can be applied for any imaging modalities, whereas the absolute method is robust for outliers although it is available for images with absolute values such as ADC. From T2WI, DWI and ADC each, 45 imaging features of morphology, histogram and texture analyses in the VOIs were extracted. The prognosis was defined based on whether recurrence within the irradiation field within two years after treatment. We constructed prediction models for locoregional recurrence with leave-one-out cross-validation using random forest algorithm, and receiver operating characteristic (ROC) analysis to evaluate diagnostic performance. definitive radiotherapy. Material and Methods

Results Cervical cancer relapsed in 25 of 107 patients within the irradiation field. The area under the curve (AUC [95% confidence interval]) calculated by ROC analysis was summarized in Table. Absolute rescaling improved AUC of ADC into 0.79 (sensitivity 69.5% and specificity 88.0% at the closest top- left point of ROC curve) on VOI2. AUCs of conventional risk factors such as volume (AUC=0.52) was significantly lower than that of ADC with radiomics (p=0.001, Delong test).

Conclusion