Venous-phase CT-based tumor habitat radiomics for differentiating retroperitoneal non-fatty dedifferentiated liposarcoma from leiomyosarcoma

To develop tumor habitat-based radiomics (H-Rad) models using multiphase computed tomography (CT) and evaluate the diagnostic value of habitat features derived from non-contrast (NP), arterial (AP), and venous-phase (VP) images for differentiating retroperitoneal non-fatty dedifferentiated liposarcoma (DDL) from leiomyosarcoma (LMS). This retrospective study included 166 patients (102 DDL, 64 LMS). Tumors were segmented on the largest axial CT slice and partitioned into three habitat subregions using Simple Linear Iterative Clustering superpixel segmentation and K-means clustering. Radiomics features were extracted from each subregion across all phases. Following LASSO regression feature selection performed independently for each subregion, XGBoost models were developed and validated for each habitat subregion across three CT phases. Performance was assessed using the area under the receiver operating characteristic curve (AUC) and SHapley Additive exPlanations (SHAP) interpretability. Habitat radiomics models derived from contrast-enhanced CT generally showed higher discriminative performance than those derived from NP images. The VP-ROI2 model achieved the best diagnostic performance, with an AUC of 0.834 (95% CI: 0.713–0.950), accuracy of 0.820, sensitivity of 0.871, and specificity of 0.737 in the testing cohort. VP-ROI2 showed nominally higher performance than NP-ROI2 ( P = 0.048), but this difference did not remain significant after FDR correction (adjusted P = 0.173). ROI2 emerged as the most discriminative habitat subregion in both AP and VP images. SHAP analysis identified original_shape_MajorAxisLength and wavelet-based texture features as the most influential predictors. Tumor habitat-based radiomics on multiphase CT enables noninvasive differentiation of non-fatty DDL from LMS by quantifying intratumoral heterogeneity. Venous-phase intermediate-attenuation subregion (ROI2) yields the highest diagnostic performance (AUC = 0.834) and may serve as a potential imaging biomarker for preoperative tumor subtyping. Further validation is required before its application in clinical decision-making.