Imaging parameters and Machine Learning models to evaluate the prognostic role of 18F-FDG PET in staging Endometrial Cancer patients

Introduction: To investigate the role of ¹⁸F-FDG PET with respect pathological prognostic factors of endometrial cancer (EC) patients, and to develop and validate machine learning (ML) models for the preoperative prediction of EC risk stratification and prognosis.

Methods: The current study included 123 female patients (median age: 66 years, range: 24-87) with histologically proven EC who underwent to ¹⁸F-FDG PET (38 PET/MRI, 85 PET/CT) at IRCCS San Raffaele Scientific Institute between August 2009 and February 2021 for staging purpose.

The following semi-quantitative PET parameters were derived on the PET positive EC primary lesions: maximum standardized uptake value (SUVmax), SUV mean, metabolic tumour volume (MTV) and total lesion glycolysis (TLG) calculated at 40% threshold of the SUVmax (SUVmean40, MTV40, TLG40).

To control for the tomograph effect, data was harmonized prior to any analysis. The nonparametric Mann–Whitney U test was used to investigate the potential of PET parameters in differentiating patients according to different pathological prognostic factors, including histological subtype (endometrioid vs. non-endometrioid EC), myometrial invasion, cancer risk group, lymph-nodal (LN) involvement, and p53 mutation. Correction for multiple testing was performed. The population was split into a train (80%) and a test (20%) set: first, significant PET parameters of the training cohort were evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC), and optimal cut-off values were derived; then, obtained cut-off values were validated using the test set. Finally, parameters showing predictive ability were used combined as inputs for the generation of multivariate ML models, precisely Random Forest Classifiers, that have been trained and validated on the same train and test sets, respectively. Test sets’ metrics scores derived from both the univariate and multivariate analyses were compared.

Results: Of the 123 EC patients, 85/123 presented an endometrioid EC, 53/115 presented myometrial invasion, 76/119 were grouped as high-intermediate / high risk group (vs. low/intermediate), 14/90 presented LN invasion, and 37/51 had p53 mutation. Ability of the PET semi-quantitative parameters in differentiating patients according to the different pathological prognostic factors are showed in Table 1, along with their AUCs evaluating the correspondent predictive value.

Optimal cut-off values derived from ROC curves were validated on the test set, and derived accuracies, sensitivities, specificities, and positive/negative predictive values were compared to the ones obtained through Random Forest Classifiers models (Table 2).

Conclusions: ¹⁸F-FDG PET parameters showed a relevant role in assessing EC pathological prognostic factors and, thus, a promising function in EC patients’ risk stratification and prognosis. Also, this study reveals that the combination of PET parameters through the generation of Machine Learning models is a valuable approach to achieve higher predictive performances.

• Download figure
• Open in new tab
• Download powerpoint