Abstract
BACKGROUND: To develop and validate an online individualized model for predicting local recurrence-free survival (LRFS) in esophageal squamous cell carcinoma (ESCC) treated by definitive chemoradiotherapy (dCRT). METHODS: ESCC patients from three hospitals were randomly stratified into the training set (715) and the internal testing set (179), and patients from the other hospital as the external testing set (120). The important radiomic features extracted from contrast-enhanced computed tomography (CECT)-based subregions clustered from the whole volume of tumor and peritumor were selected and used to construct the subregion-based radiomic signature by using COX proportional hazards model, which was compared with the tumor-based radiomic signature. The clinical model and the radiomics model combing the clinical factors and the radiomic signature were further constructed and compared, which were validated in two testing sets. RESULTS: The subresion-based radiomic signature showed better prognostic performance than the tumor-based radiomic signature (training: 0.642 vs. 0.621, internal testing: 0.657 vs. 0.638, external testing: 0.636 vs. 0.612). Although the tumor-based radiomic signature, the subregion-based radiomic signature, the tumor-based radiomics model, and the subregion-based radiomics model had better performance compared to the clinical model, only the subregion-based radiomics model showed a significant advantage (p < 0.05; training: 0.666 vs. 0.616, internal testing: 0.689 vs. 0.649, external testing: 0.642 vs. 0.604). The clinical model and the subregion-based radiomics model were visualized as the nomograms, which are available online and could interactively calculate LRFS probability. CONCLUSIONS: We established and validated a CECT-based online radiomics nomogram for predicting LRFS in ESCC received dCRT, which outperformed the clinical model and might serve as a powerful tool to facilitate individualized treatment. TRIAL REGISTRATION: This retrospective study was approved by the ethics committee (KY20222145-C-1).