Whole-tumor dual-layer spectral detector CT quantification for preoperative identification of poorly differentiated cervical squamous cell carcinoma: An interpretable prediction model

Objectives To develop an interpretable dual-layer spectral detector CT (DLCT) model for preoperative identification of poorly differentiated cervical squamous cell carcinoma (CSCC), and to compare two-dimensional region-of-interest (2D ROI) with whole-tumor three-dimensional volume-of-interest (3D VOI) parameter-extraction strategies. Methods This retrospective study included 205 surgically treated patients with CSCC who underwent contrast-enhanced DLCT before surgery. Patients were assigned to training (n = 113), internal validation (n = 47), and temporal validation (n = 45) cohorts. Parameters were extracted using 2D ROI– and 3D VOI–based analyses. Clinical, 2D DLCT, 3D DLCT, and DLCT–clinical fusion models were developed using multivariable logistic regression and evaluated by discrimination, calibration, and decision curve analysis. Results 3D VOI–based analysis showed higher interobserver reproducibility than 2D ROI–based analysis (mean intraclass correlation coefficient, 0.950 vs. 0.875; P < 0.001). The 3D DLCT model included arterial-phase 70-keV virtual monoenergetic attenuation (CT 70keV-AP ), arterial-phase normalized effective atomic number (NZeff AP ), and venous-phase normalized iodine density (NID VP ). Its areas under the receiver operating characteristic curve (AUCs) were 0.897, 0.895, and 0.885 in the training, internal validation, and temporal validation cohorts, respectively. The 2D DLCT and 3D DLCT models did not differ significantly in AUCs after multiple-comparison correction. The fusion model achieved AUCs of 0.900, 0.916, and 0.881, respectively, outperforming the clinical model without significantly improving on the 3D DLCT model. Conclusions 3D DLCT quantification provides an interpretable and noninvasive imaging biomarker for preoperative identification of poorly differentiated CSCC and may support individualized risk stratification. External validation in larger multicenter cohorts is needed to confirm generalizability.