Fig. 2: Unsupervised clustering of radiomic (3D) tumor features in 138 EC patients yields distinct clusters displaying different prevalence of high-risk features.

a Preoperative pelvic MRI with manual tumor segmentation of the primary tumor (red arrows) of a patient with endometrioid type, histologic grade 1, and FIGO stage IA. The following MRI sequences were used for radiomic profiling of the manual segmentation-cohort; contrast-enhanced volumetric interpolated breath-hold examination (VIBE + C), apparent diffusion coefficient (ADC) map and diffusion-weighted sequence with b-value of 1000 (b1000) using the segmentation mask as that for VIBE + C. b Unsupervised clustering reveals three radiomic clusters with differences in clinico-pathological variables, reflecting differences in their risk-profiles. Patients in cluster 2a (n = 44) and cluster 2b (n = 24) more often had high-risk clinico-pathological features compared to patients in cluster 1 (n = 70). c Representing feature importance of the 53 derived radiomic features in terms of pairwise cluster centroid inter-distance (solid lines with open dots) and the aggregated value (open dots, “All clusters”; scaled for visualization). A large pairwise inter-distance indicates discriminating properties of large importance between clusters for a given texture feature. d–f Kaplan–Meier survival curves depicting significantly reduced disease-specific survival among radiomic clusters 1 and 2a/b combined (d), all three clusters (e), or by recurrence (f). The number of events in brackets. Histological types; EEC endometrioid, CS carcinosarcoma, S serous, CC clear cell, U undifferentiated.