Abstract
IntroductionThis study sought to develop a predictive model using CT-based habitat radiomics to forecast pathological complete response (pCR) and progression-free survival (PFS) in esophageal squamous cell carcinoma (ESCC) patients receiving standardized neoadjuvant chemoradiotherapy (nCRT) followed by curative surgery.MethodsWe retrospectively analyzed baseline CT imaging data from 228 ESCC patients in a prospective cohort database. Patients were randomly divided into training and validation sets (7:3 ratio). Whole-tumor and habitat-derived radiomic features were extracted from pretreatment CT scans. For pCR prediction, habitat signatures were developed using Logistic Regression (LR), RandomForest (RF), and XGBoost models, optimized via grid search. PFS prediction employed Cox proportional hazards modeling with selected features. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), Hosmer-Lemeshow calibration curves, and decision curve analysis.ResultsThe habitat models retained 10 features for pCR prediction and 12 for PFS prediction. For pCR, the habitat-derived RF model demonstrated superior performance (training AUC: 0.821; validation AUC: 0.826), outperforming both other habitat models and the whole-tumor radiomics model (training AUC: 0.645). Similarly, the habitat-based RF model for PFS achieved higher AUCs (training: 0.759, 95% CI: 0.627-0.889; validation: 0.810, 95% CI: 0.653-0.966) compared to whole-tumor radiomics (training: 0.623; validation: 0.519).ConclusionOur analyses indicated a trend where habitat radiomics might outperform whole-tumor radiomics in predicting pCR and PFS for resectable ESCC after nCRT. While this merits further investigation, current evidence is insufficient to confirm its clinical utility for personalized treatment guidance.