Machine learning-based radiomics model: prognostic prediction and mechanism exploration in patients with endometrial cancer.

Objectives: To investigate the predictive value of machine-learning-based Radiomics models for postoperative overall survival (OS) of endometrial cancer (EC) patients and their biological mechanisms.

Methods: Data from 469 patients with endometrial cancer in three Centers (271 in Center 1, 154 in Center 2, and 44 in Center 3) were retrospectively and 90 patients in Center 1 were prospectively analyzed. Three-dimensional Radiomics parameters of the primary lesion and its surrounding 5 mm region in T2WI were collected from all patients. Ten machine learning methods were used to calculate the optimal Radiomics score (Radscore), whose incremental value to the available clinical indexes, pathomics, transcriptomics, and proteomics were revealed. Eventually, TCGA and CPTAC were used for the exploration of biological mechanisms of Radiomics model, with experimental validation.

Results: Radiomics features of tumor and peritumor showed some complementarity in the prognostic prediction of EC patients. The best predictive efficacy was demonstrated by the combined Radiomics model based on XGboost, with AUCs of 0.862, 0.885, 0.870 (validation set) and 0.823, 0.869, 0.849 (test set 1) and 0.850, 0.731, 0.800 (test set 2). Radiomics models demonstrated high incremental value to existing clinical indicators and can effectively improve prognostic prediction. In addition, Radiomics models have been shown to have synergistic prognostic predictive potential with pathomics, transcriptomics, and proteomics. Finally, mechanical explorations suggest that Radiomics models may be associated with tumor angiogenesis-related pathways, of which FLT1 was highlighted.

Conclusions: Machine learning-based Radiomics model contributes to predicting postoperative OS in EC patients and suggests a correlation with tumor angiogenesis.