Abstract
Objective:
Preoperatively, ovarian teratomas are often misclassified into immature and mature types, potentially resulting in adverse clinical outcomes. This study compared the diagnostic accuracy of radiomics-based machine learning (ML), deep learning (DL), and subjective radiologic assessment for differentiating immature from mature ovarian teratomas on abdominal computed tomography (CT) and evaluated whether combining these approaches improves discrimination.
Methods:
We retrospectively reviewed imaging data of 31 cases with pathologically confirmed immature teratomas and 61 age- and size-matched cases of mature teratomas (mean age=18.1±9.2 y vs. 20.9±10.0 y). Two radiologists independently annotated predefined CT features to construct a multivariable logistic regression model. Radiomics features were extracted from precontrast CT with 3 masks: whole tumor, calcification, and fat. Radiomics models were developed using 4 ML pipelines (logistic regression, random forest, support vector machine, and neural network). DL models were constructed with a 3D-ResNet18 on full-volume and tumor-segmented CT, with and without transfer learning (TL). A combined model was constructed by averaging their calibrated predicted probabilities. The diagnostic performance was assessed by the mean area under the receiver-operating characteristic curve (AUC).
Results:
In the full cohort, the 4 individual approaches performed similarly, with overlapping CIs: wavelet-transformed precontrast radiomics 0.820 (95% CI: 0.798-0.841), precontrast radiomics 0.817, subjective assessment 0.814, and DL 0.810. Combining them outperformed every individual approach (0.865; ΔAUC: 0.045, 95% CI: 0.032-0.060; Holm-adjusted
P
P
Conclusion:
Radiomics shows moderate performance in CT-based discrimination between immature and mature ovarian teratomas, with subjective radiologic assessment and DL performing at a comparable level. Combining the 3 consistently outperformed anyone alone, indicating that quantitative image analysis complements rather than replaces expert visual interpretation and may help reduce preoperative misclassification.