Role of Machine Learning (ML)-Based Classification Using Conventional (18)F-FDG PET Parameters in Predicting Postsurgical Features of Endometrial Cancer Aggressiveness

SIMPLE SUMMARY: Early and accurate assessment of endometrial cancer (EC) aggressiveness is of utmost importance for correct treatment in affected patients. However, features of EC aggressiveness are currently assessable only after surgery. The aim of the present study was to investigate the role of machine learning (ML)-based classification using (18)F-FDG PET parameters in preoperatively characterizing and predicting features of EC aggressiveness. Precisely, a signature integrating the most conventional PET parameters and clinical data was built. As a result, the described approach allowed the characterization and prediction of the investigated features of EC aggressiveness, demonstrating how advanced PET image analysis based on conventional quantitative parameters and ML can complement qualitative analysis, supporting the non-invasive preoperative stratification and treatment management of EC patients, in an interpretable and applicable way. ABSTRACT: Purpose: to investigate the preoperative role of ML-based classification using conventional (18)F-FDG PET parameters and clinical data in predicting features of EC aggressiveness. Methods: retrospective study, including 123 EC patients who underwent (18)F-FDG PET (2009–2021) for preoperative staging. Maximum standardized uptake value (SUVmax), SUVmean, metabolic tumour volume (MTV), and total lesion glycolysis (TLG) were computed on the primary tumour. Age and BMI were collected. Histotype, myometrial invasion (MI), risk group, lymph-nodal involvement (LN), and p53 expression were retrieved from histology. The population was split into a train and a validation set (80–20%). The train set was used to select relevant parameters (Mann-Whitney U test; ROC analysis) and implement ML models, while the validation set was used to test prediction abilities. Results: on the validation set, the best accuracies obtained with individual parameters and ML were: 61% (TLG) and 87% (ML) for MI; 71% (SUVmax) and 79% (ML) for risk groups; 72% (TLG) and 83% (ML) for LN; 45% (SUVmax; SUVmean) and 73% (ML) for p53 expression. Conclusions: ML-based classification using conventional (18)F-FDG PET parameters and clinical data demonstrated ability to characterize the investigated features of EC aggressiveness, providing a non-invasive way to support preoperative stratification of EC patients.