Predicting the efficacy of radiotherapy for esophageal squamous cell carcinoma based on enhanced computed tomography radiomics and combined models

{"Background": "Esophageal cancer is the eighth most prevalent and sixth most lethal cancer worldwide, with squamous cell carcinoma being the most common subtype in Asia and Eastern Europe. Over 70% of patients present at intermediate to advanced stages, and the 5-year survival rate for locally advanced disease treated with radiation is only 36\u201347%. Early identification of nonresponders to radiotherapy is critical for individualized treatment. CT-based radiomics can quantify tumor heterogeneity beyond visible morphology, but few studies have incorporated dosimetric parameters into predictive models. This study aimed to investigate enhanced CT-based radiomics and dosimetric parameters for predicting radiotherapy response in esophageal cancer.", "Methods": "A retrospective analysis included 147 patients (113 males, 34 females; median age 66 years) with esophageal squamous cell carcinoma treated between January 2018 and December 2021. Patients received 3D-CRT, IMRT, or VMAT. Treatment response was assessed at 3 months using RECIST criteria: CR and PR were classified as responders (n=89), SD and PD as nonresponders (n=58). CT images were acquired with a Siemens large-aperture scanner (120 kVp, 200 mAs, 5 mm slice thickness). The primary GTV was delineated on Oncentra software. A total of 851 radiomics features (107 original, 744 wavelet) were extracted from each GTV using 3D Slicer. Patients were split into training (n=104) and validation (n=43) groups. Feature selection used mRMR followed by LASSO with tenfold cross-validation. Logistic regression was used to construct the Rad-score. Univariate and multivariate logistic regression identified significant clinical and dosimetric predictors. A combined model incorporating Rad-score, sex, and esophageal cancer thickness was built and presented as a nomogram. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, and decision curve analysis.", "Key Results": "Seven radiomics features were selected: Original_firstorder_90Percentile, Wavelet.Hll_firstorder_Skewness, Original_glszm_HighGrayLevelZoneEmphasis, Wavelet.Lhh_glcm_ClusterShade, Wavelet.Hll_glcm_ClusterShade, Wavelet.Hll_firstorder_Maximum, and Wavelet.Hhl_glcm_ClusterShade. Univariate analysis showed sex (p=0.031) and esophageal cancer thickness (p=0.028) were significantly associated with treatment response; no dosimetric parameters were significant. In multivariate analysis, only Rad-score remained significant (OR 15.326, 95% CI 3.687\u201363.693, p<0.01). The radiomics model achieved AUCs of 0.76 (95% CI 0.67\u20130.85) and 0.73 (95% CI 0.58\u20130.88) in training and validation groups, with sensitivities of 80.5% and 88.2%, and specificities of 61.9% and 61.5%, respectively. The clinical model (sex + thickness) had AUCs of 0.65 (95% CI 0.54\u20130.76) and 0.60 (95% CI 0.42\u20130.78). The combined model showed improved performance: AUCs of 0.78 (95% CI 0.69\u20130.87) and 0.79 (95% CI 0.65\u20130.93), with sensitivities of 96.7% and 92.3%, and specificities of 54.8% and 53.3% in training and validation groups, respectively. The combined model significantly outperformed the clinical model (DeLong test, p<0.001), but not the radiomics model (p=0.772). Decision curve analysis showed the combined model had greater net benefit than the clinical model when the risk threshold exceeded 23%.", "Clinical Implications": "The combined model integrating CT radiomics features with clinical factors (sex and tumor thickness) provides a noninvasive, cost-effective tool for predicting radiotherapy response in esophageal squamous cell carcinoma. With a maximum sensitivity of 96.7%, the model may help identify patients likely to achieve CR or PR, potentially guiding treatment intensification or de-escalation. The finding that dosimetric parameters did not significantly predict response suggests that tumor heterogeneity captured by radiomics is more informative than dose-volume metrics alone. Limitations include single-institution design, small sample size, and lack of genetic feature integration. Multicenter prospective validation is needed before clinical implementation."}