Usability of synthesized image using generative adversarial network for prediction model of recurrence after radiotherapy in locally advanced cervical cancer

Abstract

PurposeTo developa generative adversarial network-based image synthesis (IS) model capable of predicting recurrence after radiotherapy in locally advanced cervical cancer. MethodsT1- and T2-weighted magnetic resonance (MR) images were synthesised using cohorts from The Cancer Imaging Archive. The results of the IS model were evaluated by comparison with real MR images. The trained IS model synthesised the MR images for the input images in the radiomics analysis dataset. A prediction model for recurrence after radiotherapy was constructed using radiomics features from real and synthesized MR images. The accuracy, specificity, sensitivity, and receiver operating characteristic curves were evaluated. ResultsUsing the LASSO regression analysis, seven and six features were extracted from the real and synthesized T1-weighted MR images, respectively, and five and seven features were extracted from the real and synthesized T2-weighted MR images, respectively. The average prediction accuracies of the five cross-validations were 78.9% and 74.3% for the real and synthesized T1-weighted MR images and 81.9% and 81.6% for the real and synthesized T2-weighted MR images, respectively. The average prediction accuracies for combined model of the real T1-weighted and real T2-weighted MR images was 90.3%, that of the real T1-weighted and synthesized T2-weighted MR images was 90.6%, and that of the real T1-weighted and synthesized T2-weighted MR images was 83.8%. ConclusionThe prediction performance of the synthesized MR image was equivalent to that of the real MR image. The prediction performance can be improved by combining the scanned and synthesized MR images.