Diagnostic model of low-grade serous ovarian cancer and borderline serous ovarian tumor: Model development and validation based on radiomics.

Abstract

e17534 Background: Evidence to date indicates that there is a close relationship between low-grade serous ovarian cancer (LGSOC) and serous borderline ovarian tumor, but there is a significant difference in prognosis. Preoperative imaging and clinical information to assess the nature of lesions as accurately as possible is beneficial to the surgical management of patients. Methods: A total of 104 patients who underwent CT examination and surgical treatment for pelvic masses in Qilu Hospital of Shandong University from 2013 to 2021 were included in the study, including 25 patients with low-grade serous ovarian cancer, 34 patients with serous borderline ovarian tumor, and 45 patients with serous cystadenoma. Quantitative image features were extracted from CT and combined with clinical information to construct a diagnostic model. The diagnostic performance was evaluated by ROC and AUC. Results: The diagnostic model was constructed by RFECV (recursive feature elimination and cross validation) together with SVM (support vector machine), random forest and XGBoost (eXtreme Gradient Boosting) algorithms, among which the RFECV combined random forest algorithm model showed the optimal accuracy (0.78) and AUC (0.92). Combined with clinical features, a new diagnostic model was constructed again by RFECV and random forest, and the accuracy (0.91) and AUC (0.96) were further improved. Conclusions: Through the combination of CT image-based radiomics and clinical features, the clinical diagnosis model was successfully constructed, which can accurately evaluate and predict the pathological types of patients, and is expected to be used in clinical practice and guide surgical decision-making.