Deep learning radiomics nomogram based on enhanced CT for differentiating epithelial ovarian borderline tumors and early-stage epithelial ovarian cancers.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiology, First Affiliated Hospital of Hebei North University, Zhangjiakou, 075000, China. [email protected].
- Hebei North University, Zhangjiakou, 075000, China.
- Department of Gynecology, First Affiliated Hospital of Hebei North University, Zhangjiakou, 075000, China.
- School of Information Science and Engineering, Hebei North University, Zhangjiakou, 075000, China.
- Department of Radiology, First Affiliated Hospital of Hebei North University, Zhangjiakou, 075000, China.
- Department of Radiology, Tianjin's Clinical Research Center for Cancer, State Key Laboratory of Druggability Evaluation and Systematic Translational Medicine, Key Laboratory of Cancer Prevention and Therapy, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, China.
- Department of Radiology, Tianjin's Clinical Research Center for Cancer, State Key Laboratory of Druggability Evaluation and Systematic Translational Medicine, Key Laboratory of Cancer Prevention and Therapy, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, China. [email protected].
Abstract
To develop an enhanced CT-based deep learning radiomics nomogram (DLRN) for differentiating between epithelial ovarian borderline tumors (EOBTs) and early-stage epithelial ovarian cancers (EOCs) and to assess its diagnostic performance using internal validation. A total of 315 patients with EOBTs or early-stage EOCs were randomly divided into a training dataset (N = 220) and a test dataset (N = 95). Six clinical factors were collected. Radiomics features were extracted from the CT images of each patient. Deep learning (DL) features were extracted using the largest axial slices of the tumor area. The clinical model was developed based on the independent clinical risk factors. Correlation analysis, univariate analysis, and least absolute shrinkage and selection operator were used to select radiomics and DL features, and the radiomics score (RADscore) and DL score (DLscore) were calculated. Radiomics and DL models were built based on RADscore and DLscore, respectively. A DLRN was constructed by integrating the independent clinical factors, RADscore and DLscore. Model performance was evaluated through receiver operating characteristic (ROC) analysis, calibration curve analysis, decision curve analysis (DCA), and DeLong test. A two-tailed P < 0.05 was considered significantly different. The DLRN exhibited satisfactory discrimination between EOBTs and early-stage EOCs with areas under the ROC curve (AUCs) of 0.887 (95% confidence interval [CI] 0.842, 0.932) and 0.900 (95% CI 0.838, 0.962) in the training and test datasets, respectively. The AUC significantly exceeded those of the clinical, radiomics and DL models in both the training and test datasets (all P < 0.05). Moreover, the DLRN achieved the lowest Brier scores among all models in both datasets. Furthermore, calibration curves exhibited good agreement between predicted and observed probabilities for the DLRN, and DCA suggested that the DLRN may provide higher net benefit than the comparison models across a range of threshold probabilities. The DLRN based on contrast-enhanced CT exhibited promising performance in differentiating between EOBTs and early-stage EOCs.