Discriminating Between Benign and Malignant Solid Ovarian Tumors Based on Clinical and Radiomic Features of MRI

Rationale and Objectives To develop and validate a combined model integrating clinical and radiomic features to non-invasive discriminate between the benign and malignant solid ovarian tumors. Materials and Methods A total of 148 patients with 156 solid ovarian tumors (86 benign and 70 malignant tumors) were included in this study. The dataset was split into the training and the test set with a ratio of 8:2 using stratified random sampling. 12 clinical features and 1612 radiomic features were extracted from each tumor. These features were selected by least absolute shrinkage and selection operator (Lasso). Three classification models were built using extreme gradient boosting (XGB) algorithm: clinical model, radiomic model, combined model. The area under the receiver operating characteristic curve (AUC), accuracy, precision and sensitivity were analyzed to evaluate the performance of these models. Results All of the three models obtained good performances in differentiating benign with malignant solid ovarian tumors in both training and test sets. The AUC, accuracy, precision, sensitivity of clinical model and radiomic model in test set were 0.847 (95% confidence interval (CI), 0.707-0.986, p <0.01), 0.774, 0.769, 0.714, and 0.807 (95%CI, 0.652-0.961, p <0.05), 0.677, 0.643, 0.643, respectively. Combined model had the best prediction results, the AUC, accuracy, precision and sensitivity were 0.954 (95%CI, 0.862-1.0, p <0.01), 0.839, 0.909 and 0.714 in test set. Conclusion Radiomics based on machine learning can be helpful for radiologists in differentiating the benign and malignant solid ovarian tumors. To develop and validate a combined model integrating clinical and radiomic features to non-invasive discriminate between the benign and malignant solid ovarian tumors. A total of 148 patients with 156 solid ovarian tumors (86 benign and 70 malignant tumors) were included in this study. The dataset was split into the training and the test set with a ratio of 8:2 using stratified random sampling. 12 clinical features and 1612 radiomic features were extracted from each tumor. These features were selected by least absolute shrinkage and selection operator (Lasso). Three classification models were built using extreme gradient boosting (XGB) algorithm: clinical model, radiomic model, combined model. The area under the receiver operating characteristic curve (AUC), accuracy, precision and sensitivity were analyzed to evaluate the performance of these models. All of the three models obtained good performances in differentiating benign with malignant solid ovarian tumors in both training and test sets. The AUC, accuracy, precision, sensitivity of clinical model and radiomic model in test set were 0.847 (95% confidence interval (CI), 0.707-0.986, p <0.01), 0.774, 0.769, 0.714, and 0.807 (95%CI, 0.652-0.961, p <0.05), 0.677, 0.643, 0.643, respectively. Combined model had the best prediction results, the AUC, accuracy, precision and sensitivity were 0.954 (95%CI, 0.862-1.0, p <0.01), 0.839, 0.909 and 0.714 in test set. Radiomics based on machine learning can be helpful for radiologists in differentiating the benign and malignant solid ovarian tumors.