Abstract
Purpose: The FIGO stage is critical to guiding the treatments of ovarian cancer (OC). This study aimed to develop a model to predict the FIGO stage of OC through machine learning algorithms using patients’ pretreatment clinical, PET/CT metabolic, and radiomics features. Methods: OC patients who underwent PET/CT scans were enrolled and divided into two cohorts according to their FIGO stage. Volumes of interest (VOI) were manually delineated to calculate PET metabolic features, and additional radiomics PET/CT features were extracted. Eleven prediction models were developed based on four feature groups. Experiments were also conducted to evaluate performance in specific subtypes like Ovarian Clear Cell Carcinoma (OCCC) and Mucinous Ovarian Cancer (MCOC). Results: 183 patients were enrolled, providing 137 features. The best model, an adaptive ensemble, achieved an AUC of 0.819. For OCCC and MCOC groups, the model achieved an AUC of 0.808. Conclusion: AI algorithms combining PET/CT metabolic and radiomics features with clinical data can significantly improve the accuracy of preoperative staging prediction.
Introduction
In ovarian cancer (OC) management, the International Federation of Gynecology and Obstetrics (FIGO) stage serves as the primary guide for clinical decision-making. It determines whether a patient is a candidate for primary cytoreductive surgery or should instead receive neoadjuvant chemotherapy. Currently, definitive staging is often only possible post-surgery, which can lead to suboptimal initial treatment strategies.
While 18F-FDG PET/CT imaging is a standard tool for assessing tumor activity, its efficacy is limited in certain subtypes. Ovarian Clear Cell Carcinoma (OCCC) and Mucinous Ovarian Cancer (MCOC) often exhibit low FDG uptake, making them difficult to stage accurately using traditional metabolic metrics. This research explores the use of machine learning to extract high-dimensional radiomics features from PET/CT scans, integrating them with clinical data to provide a more precise preoperative assessment.
Methods
The researchers developed a robust pipeline for feature extraction and model optimization:
- Feature Extraction: A total of 137 features were collected across four domains: clinical variables (e.g., age, CA125 levels), PET metabolic features (e.g., SUV, Metabolic Tumor Volume), and 110 radiomics features from both PET and CT scans describing tumor texture and shape.
- Model Development: Eleven machine learning models were evaluated. The final "adaptive ensemble" model was designed to handle ambiguous cases (prediction probabilities between 0.4 and 0.6) using a specialized voting mechanism to improve classification accuracy.
- Data Processing: The study utilized a retrospective dataset of 183 patients from Renji Hospital. Missing clinical data were addressed using the Multivariate Imputation by Chained Equations (MICE) algorithm.
Key Results
The integration of multi-modal data significantly enhanced predictive performance. The adaptive ensemble model, utilizing the full suite of clinical, metabolic, and radiomics data, achieved the highest performance with an AUC of 0.819.
As shown in the table above, while clinical data alone provided a baseline AUC of 0.771, the addition of PET and CT radiomics features provided a substantial boost. Notably, the model remained robust even for the challenging OCCC and MCOC subtypes, maintaining an AUC of 0.808. This suggests that radiomics features can capture critical tumor characteristics that are not reflected in standard metabolic uptake values.
Discussion and Conclusion
This study highlights the value of radiomics as a "secret sauce" in oncological staging. Over half of the most influential features in the final model were radiomics-based, capturing tumor heterogeneity invisible to the naked eye. Interestingly, while CA125 is a standard clinical biomarker, it was less influential than texture-based features in this AI-driven context.
Despite these promising results, several limitations remain. The study relied on a single-center retrospective cohort and manual tumor segmentation, which is labor-intensive and prone to bias. Future work should focus on multi-center validation and the implementation of automated, AI-driven segmentation to streamline the workflow for clinical use. Ultimately, this multi-modal AI approach offers a promising path toward more personalized and accurate preoperative planning for ovarian cancer patients.
Summary updated