Fig. 1: Radiogenomics approach in the current study of 866 endometrial cancer patients.

Overview of the analytical approach integrating radiomics (red), genomics (blue), and clinical/pathological data (green). Preoperative MRI from 487 endometrial cancer patients were used for primary tumor segmentation using two approaches; (i) manual segmentation by radiologists (training cohort, 138 patients) and (ii) automated segmentation (using machine learning [ML] validation cohort, 349 patients, of which 13 patients were excluded due to failed automated tumor detection). Unsupervised clustering of extracted radiomic features in the manually segmented cohort (n = 138) yielded distinct radiomic clusters tested for differences in survival and clinico-pathological characteristics. Similarly, upon feature extraction, patients assigned to clusters by the automated segmentation approach (n = 336) were tested for survival differences. Resected tumors were profiled by transcriptome expression and analyzed for molecular markers. Transcriptome profiles were obtained by L1000 and Agilent expression array data (554 patients, 98 overlappings). For a subset of patients (hexagon, n = 51), MRI, expression profiles, and TCGA molecular classes were available, enabling the generation of a gene signature. The gene signature was validated in all transcriptome datasets, including the external TCGA RNA sequencing expression dataset (n = 298) and evaluated in survival analysis.