Deep Learning Radiomics Nomogram Based on MRI for Differentiating between Borderline Ovarian Tumors and Stage I Ovarian Cancer: A Multicenter Study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, the Affiliated Suzhou Hospital of Nanjing Medical University, Suzhou Municipal Hospital, Gusu School, Nanjing Medical University, Suzhou, Jiangsu, China (X.W., X.C., Y.C., S.C., M.W.).
- Department of Radiology, The First Affiliated Hospital of Soochow University, Suzhou, Jiangsu, China (T.Q.).
- Computer Science and Engineering, University of California, Davis, Sacramento, CA (M.G.).
- Department of Radiology, The Fourth Affiliated Hospital of Soochow University, Suzhou, Jiangsu, China (Y.Z.).
- Department of Radiology, The Affiliated Huaian No. 1 People's Hospital of Nanjing Medical University, Huaian, Jiangsu, China (G.B.).
- Department of Radiology, the Affiliated Suzhou Hospital of Nanjing Medical University, Suzhou Municipal Hospital, Gusu School, Nanjing Medical University, Suzhou, Jiangsu, China (X.W., X.C., Y.C., S.C., M.W.). Electronic address: [email protected].
Abstract
To develop and validate a deep learning radiomics nomogram (DLRN) based on T2-weighted MRI to distinguish between borderline ovarian tumors (BOTs) and stage I epithelial ovarian cancer (EOC) preoperatively. This retrospective multicenter study enrolled 279 patients from three centers, divided into a training set (n = 207) and an external test set (n = 72). The intra- and peritumoral radiomics analysis was employed to develop a combined radiomics model. A deep learning model was constructed based on the largest orthogonal slices of the tumor volume, and a clinical model was constructed using independent clinical predictors. The DLRN was then constructed by integrating deep learning, intra- and peritumoral radiomics, and clinical predictors. For comparison, an original radiomics model based solely on tumor volume (excluding the peritumoral area) was also constructed. All models were validated through 10-fold cross-validation and external testing, and their predictive performance was evaluated by the area under the receiver operating characteristic curve (AUC). The DLRN demonstrated superior performance across the 10-fold cross-validation, with the highest AUC of 0.825±0.082. On the external test set, the DLRN significantly outperformed the clinical model and the original radiomics model (AUC = 0.819 vs. 0.708 and 0.670, P = 0.047 and 0.015, respectively). Furthermore, the combined radiomics model performed significantly better than the original radiomics model (AUC = 0.778 vs. 0.670, P = 0.043). The DLRN exhibited promising performance in distinguishing BOTs from stage I EOC preoperatively, thus potentially assisting clinical decision-making.