Two-staged CT-based radiomics model in characterising early-stage ovarian carcinoma and benign ovarian masses.
Authors
Affiliations (8)
Affiliations (8)
- Department of Diagnostic Radiology, School of Clinical Medicine, Li Ka Shing Faculty of Medicine, University of Hong Kong, Hong Kong, China.
- Department of Radiology, Pamela Youde Nethersole Eastern Hospital, Hong Kong, China.
- Department of Medical Imaging, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Sun Yat-sen University Cancer Centre, Guangzhou, China.
- Department of Radiology, Queen Mary Hospital, Hong Kong, China.
- Department of Radiology and Imaging, Queen Elizabeth Hospital, Hong Kong, China.
- Department of Radiology, St. Paul's Hospital, Hong Kong, China.
- Department of Pathology, School of Clinical Medicine, LKS Faculty of Medicine, University of Hong Kong, Hong Kong, China.
- Department of Diagnostic Radiology, School of Clinical Medicine, Li Ka Shing Faculty of Medicine, University of Hong Kong, Hong Kong, China. [email protected].
Abstract
Characterisation of CT detected ovarian masses is challenging with overlapping imaging features, unreliable biomarker or clinical presentation. We proposed a two-staged CT-based radiomics model to identify early-stage ovarian carcinoma (ES-OC) and sub-classify different types of benign ovarian masses (BOM). Patients with histologically confirmed BOM or ES-OC (FIGO I-II) were retrospectively recruited from 5 centres. Radiomics features were derived from CT images using PyRadiomics (v3.0.1), which intrinsically resampled volumes to isotropic 1 mm³ voxels. To reduce feature redundancy, features with high correlation (Spearman's ρ ≥ 0.85) were excluded. Two-staged feature selection was applied. First, elastic-net regression with repeated 5-fold stratified cross-validation (100 iterations) was performed to identify highly repeatable features, followed by Mann-Whitney U testing for statistical significance. Second, Boruta algorithm with Random Forest (RF) estimator was employed over 500 iterations to robustly select features by comparing their importance to randomized shadow features. Several machine learning (ML) classifiers were evaluated using stratified 10‑fold GridSearch cross-validation with area under the curve (AUC) as tuning metric. The optimal model from each stage with highest cross-validated AUC was then evaluated on the respective test set. The AUC, calibration plot, and decision curve analysis (DCA) were employed to assess the performance and clinical utility of models. The study enrolled 483 patients with 529 lesions (ES-OC: 192 patients, 192 lesions; BOM: 291 patients, 337 lesions). In the first-stage, logistic regression (LR) algorithm was selected with high sensitivity (0.870), moderate specificity (0.719) and high AUC (0.859) in the test set. In the second-stage, support vector machines (SVM) had high diagnostic accuracy with sensitivity 0.750, specificity 0.839 and AUC 0.918. DCA identified the highest benefit at 0.20 risk threshold probability in determining ES-OC. The two-staged CT-based radiomics model incorporating LR and SVM algorithms had high diagnostic efficiency in characterising ES-OC and BOM, potentially in triaging disease and personalising care.