Sub-regional radiomics combining multichannel 2-dimensional or 3-dimensional deep learning for predicting neoadjuvant chemo-immunotherapy response in esophageal squamous cell carcinoma: a multicenter study

This study aimed to develop and compare fusion models combining sub-regional radiomics with multichannel 2D and 3D DL to predict pCR in patients with LA-ESCC undergoing NACI. A total of 271 patients from three hospitals were divided into training, internal validation, and external validation cohorts. Tumor sub-regions were identified using K-means clustering based on radiomic features, and predictive features were extracted using PyRadiomics. Among all models, the DLRad1 model (radiomics + 2D DL) demonstrated the highest performance, with an AUC ranging from 0.793 to 0.910 across cohorts. Sub-region 1 features alone achieved an AUC of 0.823, while DLRad2 (radiomics + 3D DL) and other single-modality models showed lower AUCs (0.701-0.906). Spearman correlation analysis confirmed low redundancy among selected features. These findings support DLRad1 as a promising non-invasive tool to identify LA-ESCC patients most likely to benefit from NACI, potentially aiding personalized treatment decisions.