Multiparametric MRI Radiomics and MAID-CDF Texture Descriptors for Preoperative Three-Class Histopathological Grading of Endometrial Carcinoma

Background/Objectives: To evaluate a multiparametric MRI machine-learning framework for preoperative three-class histopathological grading of endometrial carcinoma using conventional radiomics, Multiradius Annular Intensity-Difference cumulative-distribution-function descriptors (MAID-CDF), stability-based feature selection, and decision-level multimodal fusion. Methods: This retrospective single-center study included 96 patients (Grade 1, n=26; Grade 2, n=40; Grade 3, n=30). Quantitative features were derived from contrast-enhanced T1-weighted MRI (T1+C), T2-weighted MRI, and ADC maps. Model development used stratified five-fold outer cross-validation, with Elastic-Net stability selection and hyperparameter optimization performed within each training partition before support vector machine classification. Multimodal predictions were combined by hard voting, soft voting, and cross-fitted Random-Forest stacking. Performance was calculated from pooled out-of-fold predictions. The primary models were also assessed using ten independently seeded repetitions of stratified five-fold nested cross-validation. Results: T1+C, T2, and ADC models achieved accuracies of 88.5%, 89.6%, and 92.7% and macro-F1 scores of 0.884, 0.891, and 0.925, respectively. ADC had the highest single-modality macro-AUC (0.971). Hard and soft voting each achieved 93.8% accuracy and a macro-F1 of 0.937; soft voting yielded a macro-AUC of 0.989. The Random-Forest stack achieved 90.6% accuracy, a macro-F1 of 0.904, and a macro-AUC of 0.985. For ADC, adding MAID-CDF to radiomics increased Grade 3 sensitivity from 80.0% to 93.3%, although a paired exact test did not establish a significant difference for Grade 3 sensitivity. Across ten repetitions, soft voting achieved 93.0% ± 1.1% accuracy and macro-F1 of 0.929 ± 0.012; the combined ADC model achieved 90.2% ± 2.2% accuracy and macro-F1 of 0.897 ± 0.022. Conclusions: The framework achieved high internally validated performance for three-class grading. ADC provided the strongest single-modality results, while the contribution of MAID-CDF was modality dependent and appeared most favorable with ADC radiomics. These findings support further evaluation of complementary quantitative MRI representations for preoperative grading. Repeated partitioning supported the relative consistency of soft voting, whereas the additional contribution of MAID-CDF varied across performance measures.