CT deep learning radiomics and genomics for predicting staging of epithelial ovarian cancer.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China.
- Department of Radiology, Intelligent Medical Imaging of Jiangxi Key Laboratory, Nanchang, China.
- Department of Radiology, Jiangxi Provincial People's Hospital, Nanchang, China.
- Clinical and Technical Support, Philips Healthcare, Shanghai, China.
- Department of Radiology, The First Affiliated Hospital, Hengyang Medical School, University of South China, Hengyang, China.
Abstract
Accurate preoperative staging is critical for improving the prognosis, treatment, and survival outcomes of epithelial ovarian cancer (EOC) patients. To develop and validate a novel model, associating computed tomography (CT)-based radiomics features, deep learning (DL) features, and transcriptomics data with the tumor microenvironment, for predicting staging in patients with EOC. A total of 201 EOC patients from three hospitals were randomly divided into a training set (n =Â 160) and an internal validation set (n =Â 41) using an 8:2 ratio. Additionally, an external validation set comprised 84 EOC patients from The Cancer Genome Atlas. From the CT images, 1130 radiomic features and 512 DL features were extracted. The features were selected using maximum relevance minimum redundancy, followed by refining using least absolute shrinkage and selection operator (LASSO) regression. Five models were constructed via logistic regression, namely clinical-semantic (CS) model, radiomics model, DL model, CS-radiomics model, and combined model. The predictive performance of the models was assessed using area under the curve (AUC), calibration curve, and decision curve analysis (DCA). RNA sequencing data from The Cancer Image Archive were used to investigate the radiogenomics-related immune infiltration patterns. Five radiomics features and four DL features were selected by LASSO algorithm for building the radiomics and DL scores. The combined model demonstrated excellent predictive performance with AUCs of 0.910 in the training set, 0.913 in the internal validation set, and 0.882 in the external test set. DCA indicated that the combined model had good clinical applicability when the threshold probability exceeded 20%. The calibration curves of the combined model showed a good consistency between the observed value and predicted value in the training and validation sets. Furthermore, the advanced-stage group displayed a higher levels of immune infiltration compared to the early-stage group. The combined model can successfully predict staging in patients with EOC, providing reliable basis to develop personalized treatment strategies for EOC patients.