Commentary on “Intratumoral and peritumoral heterogeneity based on CT to predict the pathological response after neoadjuvant chemoimmunotherapy in esophageal squamous cell carcinoma”

Dear Editor, We read with great interest the article by Ling et al, entitled “Intratumoral and peritumoral heterogeneity based on CT to predict the pathological response after neoadjuvant chemoimmunotherapy in esophageal squamous cell carcinoma”[1]. By incorporating clinical variables along with intratumoral and peritumoral heterogeneity information, this study aims to achieve precise prediction of major pathological response (MPR) and even pathological complete response (pCR), demonstrating significant clinical relevance and methodological innovation. The comprehensive study design, combining retrospective and prospective cohorts with detailed feature engineering, yielded a high-performing model (AUC: 0.915). Despite these strengths, several methodological and clinical translation issues require further discussion to optimize and facilitate real-world applicability. First, study design and sample considerations. The study relied on data from a single medical center. Although a prospective validation cohort was included, the single-center design remains limited by the patient population characteristics, imaging acquisition protocols, pathological evaluation standards, and treatment preferences specific to that center. The cohort had an overwhelmingly high proportion of males (98.1%), which, while partially consistent with the higher prevalence of esophageal squamous cell carcinoma in males, may limit the model’s applicability to female patients. Moreover, the model’s performance may be overfitted to male-dominated population characteristics. External validation using independent multi-center datasets with more balanced demographic compositions is essential to o assess generalizability. The total sample size was 157 cases (110 in the training set and 47 in the test set), with 60 MPR-positive cases (38.2%) and 40 pCR-positive cases (25.5%). When constructing complex machine learning models that integrate high-dimensional clinical, intratumoral, and peritumoral features, sample size, particularly the number of positive events, is critical. Although sample size estimation was performed, the test set contained only 47 cases, with approximately 18 MPR-positive cases. This limited sample can result in wide confidence intervals and unstable performance estimates[2]. For instance, the test set AUC was 0.915 (95% CI: 0.844–0.981), and the wide confidence interval reflects uncertainty due to the limited sample. For endpoints with even lower positive rates, such as pCR, insufficient sample size may further affect reliability. Hence, larger-scale, independent multi-center validation is necessary to confirm the model’s true performance. Second, manual segmentation reproducibility and efficiency bottlenecks. Tumor and peritumoral regions of interest (ROIs) were manually delineated by two radiologists, with discrepancies resolved by consensus. However, manual segmentation is subjective, time-consuming, and challenging to standardize, limiting feasibility for large-scale or multicenter clinical use. Although the authors acknowledged this limitation, inter- and intra-observer agreement metrics (e.g., Dice coefficient) were not reported, which affects the reproducibility of feature extraction. Future work should prioritize automated or semi-automated segmentation methods, such as deep learning-based semantic segmentation models, to improve efficiency, reduce human variability, and facilitate integration into clinical workflows. Third, limited description of technical details. Habitat radiomics, used to characterize spatial heterogeneity, is a notable strength of the study. This approach reclassifies voxels into five habitat types based on local spatial autocorrelation and regional attribute values[3,4]. However, technical details were relatively brief. For example, the radiomics feature processing pipeline (e.g., use of raw CT values or filtered data), neighborhood weight matrix parameters, and significance thresholds were not fully described. These parameters directly influence habitat classification and subsequent feature extraction, warranting clarification. Additionally, evaluating the impact of alternative habitat classification strategies (e.g., varying category numbers or classification methods) could enhance methodological robustness and interpretative. Ling et al demonstrate that integrating peritumoral and clinical data with habitatomics can effectively predict neoadjuvant chemoimmunotherapy response in esophageal squamous cell carcinoma. Their model showed strong performance, supported by SHAP analysis and an online calculator to improve interpretability. Future work should focus on method transparency, automated segmentation, improved visualization, and prospective trials to assess real-world utility. With continued refinement, AI-radiomics models such as this could facilitate personalized treatment in esophageal cancer, in line with the TITAN guideline advocating transparency in AI reporting[5].