Combination Analysis of a Radiomics-Based Predictive Model With Clinical Indicators for the Preoperative Assessment of Histological Grade in Endometrial Carcinoma

Abstract

Background: Histological grade is one of the most important prognostic factors of endometrial carcinoma (EC) and when selecting preoperative treatment methods, conducting accurate preoperative grading is of great significance. Purpose: To develop a magnetic resonance imaging (MRI) radiomics-based nomogram for discriminating histological grades 1 and 2 (G1 and G2) from grade 3 (G3) EC. Methods: This was a retrospective study included 358 patients with histologically graded EC, stratified as 250 patients in a training cohort and 108 patients in a test cohort. T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI) and a dynamic contrast-enhanced three-dimensional volumetric interpolated breath-hold examination (3D-VIBE) were performed via 1.5-Tesla MRI. To establish ModelADC, the region of interest was manually outlined on the EC in an apparent diffusion coefficient (ADC) map. To establish the radiomic model (ModelR), EC was manually segmented by two independent radiologists and radiomic features were extracted. The Radscore was calculated based on the least absolute shrinkage and selection operator regression. We combined the Radscore with carbohydrate antigen 125 (CA125) and body mass index (BMI) to construct a mixed model (ModelM) and develop the predictive nomogram. Receiver operator characteristic (ROC) and calibration curves were assessed to verify the prediction ability and the degree of consistency, respectively. Results: All three models showed some amount of predictive ability. Using ADC alone to predict the histological risk of EC was limited in both the cohort [area under the curve (AUC), 0.715; 95% confidence interval (CI), 0.6509–0.7792] and test cohorts (AUC, 0.621; 95% CI, 0.515–0.726). In comparison with ModelADC, the discrimination ability of ModelR showed improvement (Delong test, P < 0.0001 for both the training and test cohorts). ModelM, established based on the combination of radiomic and clinical indicators, showed the best level of predictive ability in both the training (AUC, 0.925; 95% CI, 0.898–0.951) and test cohorts (AUC, 0.915; 95% CI, 0.863–0.968). Calibration curves suggested a good fit for probability (Hosmer–Lemeshow test, P = 0.673 and P = 0.804 for the training and test cohorts, respectively). Conclusion: The described radiomics-based nomogram can be used to predict EC histological classification preoperatively.