Predicting vital retroperitoneal residual tumors of metastatic testicular tumor patients after chemotherapy using radiomics.

Abstract

527 Background: About 50% of patients undergoing post-chemotherapy retroperitoneal lymph node dissection (pcRPLND) are overtreated due to missing markers or valid prediction scores prior surgery. The potential of radiomics and machine learning applied on computed tomography (CT) imaging to predict the presence of viable tumor or teratoma in retroperitoneal lymph node metastases from germ cell tumor (GCT) patients prior to pcRPLND has not been explored. Therefore, we applied radiomics and machine learning to CT images of GCT patients prior to pcRPLND. Methods: Metastasized GCT patients who were treated with chemotherapy and received a contrast-enhanced CT prior to pcRPLND possessing complete clinical data were included in the study. Only lymph nodes which were identified in the CT images and correlated with the pathological findings (benign: necrosis / fibrosis vs. viable: viable tumor/ teratoma) were included. Lymph nodes identified in the CT images, were semiautomatically segmented and 93 radiographic features were analyzed. A linear support vector machine (SVM) algorithm was applied to analyze reproducible radiomics features. Additionally, a continuous reduction of the features analyzed was performed using Random Forest algorithms, as well as consecutive correlation and receiver operating curve analyzes. Results: Forty-two patients fulfilled the inclusion criteria and were included in the study. Total in these patients 96 lymph nodes were segmented on CT. Histologically, 41 lymph nodes were classified as viable tumors and 55 as benign. To train the SVM, 67 lymph nodes were randomly selected. Of the 93 radiomic features analyzed, 51 features were reproducible. Applying the trained algorithm to the remaining 29 lymph nodes resulted in a classification accuracy of 82% with a diagnostic sensitivity of 81% and a specificity of 83%. After multistep feature reduction, the three most important predictors for viable tumor achieved a sensitivity of 66% and a specificity of 78% when combined in a multivariate model. Conclusions: The applied radiomics model, solely based on CT images achieved a good sensitivity and specificity in predicting viable metastases.