Multiparametric MRI Habitat Imaging for Preoperative Assessment of Ki-67 Proliferation Index in Meningiomas: A Multicenter Study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, The Second Affiliated Hospital of Hainan Medical University, Haikou, Hainan, China.
- Department of Nuclear Medicine, the First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi, China.
- Department of Radiology, Xiangya Hospital Central South University, Changsha, Hunan, China.
- Department of Radiology, The First Affiliated Hospital of Hainan Medical University, Haikou, Hainan, China.
- Department of Nuclear Medicine, Hainan Cancer Hospital, Haikou, Hainan, China.
- Department of Ultrasound Medicine (South Campus), Honghui Hospital, Xi'an Jiaotong University, Xi'an, Shaanxi, China.
Abstract
Preoperative assessment of meningioma proliferative activity relies on the postoperative Ki-67. Habitat imaging captures proliferative variation invisible to whole-tumor analysis by segmenting tumors into distinct subregions. To develop and validate a multiparametric MRI habitat imaging model for preoperative assessment of the Ki-67 proliferation index in meningiomas. Retrospective, multicenter. Five hundred and twelve-patients (mean age 54.4 ± 10.8 years; 340 [66.4%] female) from four institutions, divided into a Training Set (n = 220) and two independent external validation sets (n = 95 and 197); Ki-67 ≥ 5% defined the high-expression group. 1.5-T or 3.0-T; T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), contrast-enhanced T1-weighted (CE-T1), and T2-fluid-attenuated inversion recovery (T2-FLAIR) sequences (spin echo, fast spin echo, and inversion recovery sequences). K-means clustering defined four habitats. Radiomic features were extracted to construct a Bagging-multilayer perceptron (MLP) ensemble, evaluated by receiver operating characteristic (ROC) analysis, sensitivity, specificity, decision curve analysis (DCA), and SHapley Additive exPlanations (SHAP). Area under the ROC curve (AUC) with 95% confidence intervals (CIs); sensitivity, specificity, F1 score, positive predictive value (PPV), negative predictive value (NPV); DeLong tests; net reclassification improvement (NRI); integrated discrimination improvement (IDI); Brier scores; Hosmer-Lemeshow test. Two-tailed p < 0.05. Habitat-derived features, particularly textural heterogeneity in the strongly enhancing subregion, accounted for most selected features (9 of 15). The Final model achieved AUCs of 0.929 (95% CI: 0.894-0.964; Training Set), 0.903 (95% CI: 0.846-0.961; External Validation Set 1), and 0.914 (95% CI: 0.872-0.956; External Validation Set 2), significantly higher than all baseline models (ΔAUC 0.091-0.266; DeLong p < 0.05). Across cohorts, sensitivities ranged 0.651-0.843, specificities 0.874-0.911, PPVs 0.789-0.848, NPVs 0.758-0.912, and F1 scores 0.738-0.815. Compared with whole-tumor radiomics, NRI was 0.71-0.83 and IDI 0.26-0.39. DCA showed the highest standardized net benefit (0.230-0.274) at 15%-35%. Brier scores (0.106-0.152) were lower than those of the Habitat model; Hosmer-Lemeshow p values were 0.43-0.76. This habitat-based model may enable noninvasive preoperative assessment of the Ki-67 proliferation index in meningiomas and assist preoperative risk stratification; prospective validation is warranted. 2. Stage 3.