Incremental diagnostic value of tumor habitat radiomics for risk stratification in thymic epithelial tumors

To determine the incremental diagnostic value of habitat radiomics for risk stratification of thymic epithelial tumors (TETs) based on contrast-enhanced CT (CECT).

This retrospective study included 220 patients with pathologically confirmed TETs (82 high-risk [B2/B3/thymic carcinoma] and 138 low-risk [A/AB/B1]) who underwent preoperative CECT. Tumors were segmented into 3 subregions (habitats) using k-means clustering, and radiomic features were extracted from both whole-tumor and subregions. After feature selection (variance threshold, reproducibility evaluation, XGBoost-based importance ranking, and recursive feature elimination), three machine learning models were developed (1): a conventional radiomics model (2), a habitat radiomics model, and (3) a combined model integrating both feature sets. Model performance was evaluated using ROC analysis, net reclassification improvement (NRI), integrated discrimination improvement (IDI), calibration metrics, and decision curve analysis (DCA).

The combined model demonstrated superior discrimination (AUC: 0.900) compared to the conventional (AUC: 0.819) and habitat (AUC: 0.734) radiomics models in the independent test set. Although DeLong's test showed no statistically significant difference (p=0.161), the performance of combined model demonstrated incremental diagnostic value (NRI: 0.286; IDI: 0.209). Calibration and DCA confirmed its robustness and higher net benefit across decision thresholds. While the models' training performance might suggest overfitting, their test results demonstrate generalizability.

The habitat radiomics approach enables accurate risk stratification prediction in TETs and demonstrates potential as a clinically valuable tool to augment the performance of conventional radiomics models in routine practice.