Predicting aggressive disease and poor outcome in endometrial cancer using preoperative [¹⁸F]FDG PET primary tumor...

To develop a [¹⁸F]fluorodeoxyglucose ([¹⁸F]FDG) positron emission tomography (PET) primary tumor radiomic model for predicting disease-specific survival (DSS), and compare it with conventional PET markers in a large endometrial cancer cohort.

Radiomic features were extracted from preoperative [¹⁸F]FDG PET scans of 489 endometrial cancer patients using a standardized uptake value (SUV) threshold > 2.5 to define primary metabolic tumor volumes (MTVs). A second reader extracted features in 154/489 patients, in which intraclass correlation coefficients (ICCs) were calculated. Radiomic features with ICCs > 0.75 were retained and ComBat harmonization was applied to reduce scanner/protocol effects on the extracted features. Patients were divided into training (n = 343) and test (n = 146) sets. A radiomic DSS score (R_dss) was developed in the training set using least absolute shrinkage and selection operator (LASSO) Cox regression. A combined model (C_dss), incorporating R_dss, PET positive lymph nodes (LN_PET) and preoperative histology risk was constructed using multivariable Cox hazard analyses. Prediction performances were assessed by comparing areas under time-dependent receiver operating characteristic curves (tdROCs AUCs) for R_dss, C_dss, and conventional PET markers: SUV_max, SUV_mean, MTV, tumor lesion glycolysis (TLG) and LN_PET.

In the test set, AUCs for 2- and 5-year DSS were higher for R_dss (0.855, 0.720) compared to SUV_max (0.548, 0.572) and SUV_mean (0.549, 0.554) (p ≤ 0.04 for all), while similar to MTV (0.863, 0.696), TLG (0.814, 0.672) and LN_PET (0.802, 0.626) (p ≥ 0.12 for all). C_dss predicted 2-year DSS with AUC of 0.909 in the test set, outperforming all conventional imaging markers (p ≤ 0.04 for all) except MTV (p = 0.29). For 5-year DSS, C_dss (AUC: 0.817) outperformed all conventional imaging markers, including MTV (AUC ≤ 0.696, p ≤ 0.05, for all).

R_dss predicts short-term survival with high accuracy, outperforming tumor SUV_max/mean, but not MTV, TLG and LN_PET. The combined C_dss model yields high accuracy for predicting both short- and long-term survival, outperforming all conventional PET imaging markers.