Component-based CT radiomics for preoperative differentiation of well-differentiated and dedifferentiated retroperitoneal liposarcoma.

Introduction: Preoperative distinction between well-differentiated liposarcoma (WDLPS) and dedifferentiated liposarcoma (DDLPS) remains difficult on conventional imaging. We evaluated whether CT attenuation-defined regional radiomics provides additional discriminatory information beyond whole-tumor radiomics and simple quantitative CT measurements. Methods: This retrospective study included 97 patients with surgically confirmed retroperitoneal liposarcoma (57 WDLPS and 40 DDLPS) who underwent preoperative contrast-enhanced CT. Radiomics features were extracted from the whole tumor (W), CT-defined fat-attenuation (F) and soft-tissue-attenuation (S) subregions, and a 3-mm peritumoral region (R). Six prespecified feature sets (W, F, S, WF, WFS, and WFR) were modeled with L1-regularized logistic regression using nested cross-validation. Clinical and quantitative CT benchmark models were also evaluated, and the WF feature set was further compared across logistic regression, support vector machine, random forest, and XGBoost classifiers. Performance was assessed from held-out out-of-fold predictions. Results: The WF model had the highest pooled out-of-fold AUC among the six regional models (0.864; 95% CI, 0.778-0.940), but adding S or R features did not improve discrimination. The quantitative CT and clinical models achieved AUCs of 0.849 and 0.562, respectively. In a separate 10-times repeated nested cross-validation analysis, the mean fold-wise AUC difference between WF and quantitative CT was 0.023 (95% CI, -0.030 to 0.076; Holm-adjusted P = 0.389), and no pairwise comparison among the six regional models remained significant after Holm correction. When feature-set and classifier selection were repeated within the outer cross-validation loop, the selection-aware AUC was 0.842 (95% CI, 0.748-0.926). Logistic regression had the highest AUC among the four classifiers evaluated with the WF feature set. Whole-tumor mean intensity, fat-subregion run-length non-uniformity normalized, and whole-tumor GLCM correlation were the leading features in the descriptive interpretation analysis. Discussion: CT attenuation-defined regional radiomics was feasible for preoperative WDLPS-DDLPS differentiation, but the WF model did not show a significant advantage over whole-tumor radiomics or segmentation-derived quantitative CT. External validation is needed.