CT-based radiomics analysis for the prediction of response to neoadjuvant chemotherapy in esophageal cancer: a dual-center study

Abstract Introduction Esophageal cancer (EC) ranks sixth globally in cancer-related mortality, with locally advanced EC (LAEC) patients often receiving neoadjuvant chemotherapy (NACT). Approximately 40% exhibit poor treatment response (PR), leading to unnecessary toxicity and delayed interventions. Current histopathological evaluation occurs post-surgery, highlighting the need for non-invasive early prediction tools. Methods This dual-center retrospective study enrolled 268 LAEC patients (Center I: 214; Center II: 54), stratified into training (TC, n = 151), internal validation (IVC, n = 63), and external validation (EVC, n = 54) cohorts. Inclusion required: (a) pathologically confirmed LAEC (cT2-4aNxM0); (b) baseline contrast-enhanced CT; (c) NACT followed by surgery. Portal venous-phase CT images were segmented manually using MITK software. A total of 1,409 radiomics features were extracted via PyRadiomics (IBSI-compliant). These comprised 107 features derived from the original and conventionally filtered images, and 1,302 features obtained from wavelet-transformed and Laplacian-of-Gaussian (LoG)-filtered images. Feature selection included: (i) intraclass correlation coefficient (ICC > 0.8) for reproducibility; (ii) minimum redundancy maximum relevance (MRMR) ranking; (iii) Boruta algorithm finalization. Four key wavelet/LoG-filtered features were used to train random forest (RF), logistic regression (LR), and support vector machine (SVM) classifiers. Results Four robust radiomics features—DependenceVariance, LargeAreaLowGrayLevelEmphasis, LargeDependenceEmphasis, and ShortRunLowGrayLevelEmphasis—derived from wavelet transforms and Laplacian of Gaussian filters, demonstrated significant discriminative power between poor responders (PR) and non-poor responders (nPR) in the training cohort ( p < 0.05). When applied to predictive modeling, the Random Forest (RF) classifier consistently outperformed other models across all validation cohorts. Specifically, RF achieved an area under the curve (AUC) of 0.823 (95% CI: 0.757–0.889) with 75.5% accuracy in the training cohort, maintained strong performance in internal validation (AUC = 0.807, 95% CI: 0.700–0.914), and demonstrated generalizability in external validation (AUC = 0.774, 95% CI: 0.649–0.899). Statistical comparison via Delong tests confirmed RF’s significant superiority over both logistic regression and support vector machine classifiers (p < 0.05), as evidenced by receiver operating characteristic curves. This performance consistency indicates that, among patients who completed paclitaxel/cisplatin-based NACT and subsequently underwent oesophagectomy, the model identified those likely to have a poor pathological response before treatment initiation. Conclusions The RF-based radiomics model using wavelet/LoG features predicted poor pathological response to paclitaxel/cisplatin-based NACT in patients with LAEC who subsequently underwent oesophagectomy, with consistent performance across the training, internal-validation and external-validation cohorts. Because only surgically resected patients were studied, this conclusion applies to that population, and prospective validation in intention-to-treat cohorts is required before the model could be used to change management. Manual tumour segmentation remains a limitation, and future integration of automated segmentation and multi-omics data could enhance clinical utility.