objective
To develop and validate a radiomics-based machine learning nomogram using multiparametric MRI to preoperatively predict aggressive histology in endometrial cancer (EC) patients.
method
This bicenter study retrospectively analyzed patients with histologically confirmed EC who underwent preoperative MRI. Radiomics features were trained and tested to predict aggressive histology using a support vector machine (SVM) algorithm. Clinical data and conventional MRI findings were collected. Multivariable logistic regression analysis was performed to create a predictive fusion model. The model was presented as a nomogram on the training set and validated on an independent external test set. Calibration curves and the Hosmer-Lemeshow test were used to assess goodness of fit. Three predictive models were constructed: M1 (original biopsy only), M2 (radiomics only), and M3 (combined nomogram). Model performance was evaluated using ROC analysis, and pairwise comparisons of AUC were performed by DeLong’s test. DCA was used to compare net income.
result
283 women were enrolled (training: 198, testing: 85). M3 achieved an AUC of 0.900 (95% CI: 0.850 to 0.938) and 0.890 (95% CI: 0.803 to 0.948) on the training and test sets, respectively, and showed good fit according to the Hosmer-Lemeshow test (P> 0.05). Delong test with Bonferroni correction showed that the AUC of M3 of the fusion model exceeded the AUC of M1 and predicted progressive histology (adjusted) P< 0.05). Additionally, DCA demonstrated a higher net profit for the M3 model with an IDI of 0.126 and 0.176 (P< 0.01) in both sets.
conclusion
A multiparametric MRI-based radionics machine learning nomogram improves preoperative diagnosis of aggressive histology in EC patients.
