Abstract / Summary
To develop a multiparametric MRI-based fusion model integrating habitat radiomics with a Vision Transformer (ViT) to preoperatively predict p53 status in endometrial cancer (EC), and to explore its potential prognostic relevance. From three centers, this retrospective study analyzed 1,230 patients with EC, assigned to a training cohort ( n = 447), an internal validation cohort ( n = 192), and two external validation cohorts ( n = 246 and 345). We used the tumor and habitat regions to extract radiomics features. Deep learning (DL) features were extracted using the ViT model. Clinical, radiomics, habitat, ViT, and fusion models were constructed based on the support vector machine algorithm. Interpretability was enhanced using SHapley Additive Explanations (SHAP) analysis, and prognostic value for progression-free survival (PFS) was further evaluated. p53-abnormal (p53abn) immunophenotype was identified in 232 patients (18.9%). The maximum tumor diameter and histologic grade were selected as clinical predictors of p53abn status. The habitat model attained areas under the curve (AUCs) of 0.770–0.802 in the validation cohorts, outperforming the radiomics model. A fusion model integrating the habitat radiomics and DL features with clinical predictors achieved the optimal performance, with AUCs of 0.881, 0.869, and 0.904 in the internal and two external validation cohorts, respectively. SHAP analysis indicated that the habitat and DL features made significant contributions to p53abn EC prediction. Exploratory survival analysis showed that fusion-model–predicted p53 status was independently associated with PFS. The fusion model may serve as a complementary imaging tool for preoperative assessment of p53 status in EC.