fig3

Figure 3. Exploratory representation comparison. (A) Internal holdout AUROC for whole-tumor (0.690), whole-tumor plus habitat (0.710), habitat-only (0.738), and patient-balanced whole-tumor plus habitat (0.833; 95%CI: 0.683-0.955) models, displayed in representation order. Equal marker sizes and a shared color avoid emphasizing a winning model. Intervals were unavailable for the three comparators; (B) training CV mean vs. holdout AUROC; horizontal bars show CV standard deviations. The patient-balanced model had training CV AUROC 0.6915 ± 0.0887. Marker shapes identify the same models in both panels: circle, whole-tumor; square, whole-tumor plus habitat; triangle, habitat-only; diamond, patient-balanced combination. Habitat-only CV results were unavailable. Both panels use AUROC axes from 0.50 to 1.00, and the dashed diagonal denotes equal training and holdout estimates. All models were evaluated on the same reused 33-patient holdout; the comparisons are descriptive, without multiplicity correction or independent model validation. AUROC: Area under the receiver operating characteristic curve; CI: confidence interval; CV: cross-validation; MR_OTHER and T1_VIBE_FS: filename-derived MR acquisition-category identifiers, not verified imaging protocols or contrast phases; n: number of patients.