Two-Stage CT-Based Radiomic Signature to Predict the Efficacy of Immunotherapy.

PURPOSE/OBJECTIVE(S) Immunotherapy is not effective for all patients, and its prediction method still needs further research and improvement. We try to use Radiomics to predict short-term efficacy, including pseudoprogression. MATERIALS/METHODS The study retrospectively enrolled 342 CT scans from 171 NSCLC patients who received pembrolizumab or nivolumab with or without ipilimumab from June 2016 to May 2019. Two-stage CT images, including CT before and after the first treatment, were employed to develop and validate the radiomics model for predicting the efficacy of immunotherapy. There were two patients with complete response (CR), 74 patients with partial response (PR), 55 patients with stable disease (SD), 32 patients with immune progressive disease (PD), and 7 patients with immune unconfirmed progressive disease (iuPD) according to RECIST 1.1 criteria. In this study, patients with CR, PR, and iuPD were considered as one group, termed Clinical Respond (CP). We built a radiomic model to distinguish PD from CP and SD patients. Radiomic features were extracted separately from two-stage CT images. The two-stage feature errors Δ_feature = Feature_before-Feature_after were input into a random forest classifier to build the radiomic model. Five-fold cross-validation was randomly performed five times. The predictive performance was evaluated using overall accuracy for this triple classification task. Furthermore, we also used the one-vs-rest strategy, treating each class as a positive in turn and the rest as negative, to evaluate the performance with respect to the accuracy, sensitivity, and specificity. RESULTS Our proposed method achieves the performance with an overall accuracy of 64.41 ± 1.57%. The predictive model using the pre-treatment and post-treatment CT features achieves an overall accuracy of 42.92 ± 2.14% and 42.98 ± 1.51%, respectively. The predictive model using the pre-treatment and post-treatment CT features achieved 42.92 ± 2.14% and 42.98 ± 1.51%, respectively. Therefore, a predictive model using the two-stage feature errors is significantly better than using pre-treatment and post-treatment CT features alone. In addition, the other metrics of our proposed method are as follows. The CP, SD, and PD accuracy are 77.57 ± 1.03, 68.40 ± 1.07%, and 82.85 ± 1.00%, respectively. The sensitivity for CP, SD, and PD are 77.57 ± 1.03%, 68.40 ± 1.37%, and 82.86 ± 1.00%, respectively. The specificity for CP, SD, and PD are 81.11 ± 1.60%, 73.61 ± 1.83, and 90.02 ± 0.52%, respectively. CONCLUSION Radiomics can ably assist in predicting the treatment response of immunotherapy, and it can be a useful supplement in the part of the evaluation of iRecist. It is worthy of further research on imaging with larger sample size and more modalities.