Integration of habitat radiomics and traditional radiomic features for predicting pathological complete response in esophageal squamous cell carcinoma following neoadjuvant immunotherapy and chemotherapy: a multicenter comparative study

TL;DR

This study compares habitat radiomics and traditional radiomic features to predict treatment response in esophageal cancer patients receiving immunotherapy and chemotherapy.

Contribution

The study introduces a combined model of habitat and traditional radiomics that outperforms individual models in predicting pathological complete response.

Findings

Habitat radiomics model showed higher AUC and better sensitivity and specificity compared to traditional radiomics.

The combined model achieved the highest AUC of 0.960 across multiple validation cohorts.

The model demonstrates potential for personalized treatment strategies in ESCC patients.

Abstract

Esophageal squamous cell carcinoma (ESCC) remains one of the leading causes of cancer-related mortality worldwide. Although immunotherapy has shown promising efficacy for locally advanced ESCC, the lack of reliable predictive tools and the marked heterogeneity of tumors make it difficult to accurately evaluate treatment responses. To address this challenge, we conducted a multicenter study aimed at developing and comparing predictive models based on habitat radiomics and traditional radiomic features to estimate pathological complete response (pCR) in patients receiving neoadjuvant immunotherapy and chemotherapy. Using multicenter data, we systematically assessed the performance of these models to determine the relative advantages of each feature type in predicting treatment outcomes and supporting personalized therapeutic strategies.

This retrospective study analyzed ESCC patient data from three medical centers. Pre-treatment CT imaging was utilized for tumor region segmentation and the extraction of both Habitat Radiomics and traditional Radiomic Features. Feature selection was performed using LASSO regression, and machine learning models were developed based on these features. Several machine learning algorithms, including Support Vector Machines (SVM), Random Forest, and XGBoost, were employed for training and validation. Model performance was evaluated using metrics such as ROC curves, AUC, sensitivity, and specificity.

The Habitat Radiomics model achieved AUCs of 0.938 in the training cohort, 0.896 in the internal validation cohort, 0.819 in external validation cohort 1, and 0.846 in external validation cohort 2, demonstrating strong and consistent predictive performance. In comparison, the traditional Radiomics model yielded AUCs of 0.941, 0.845, 0.796, and 0.729, respectively. Beyond higher AUC values, the Habitat Radiomics model also showed superior sensitivity and specificity in predicting pCR. Notably, the combined model that integrated both Habitat and traditional Radiomics features outperformed the individual models, achieving the highest AUC of 0.960 across cohorts and underscoring its superior predictive accuracy.

This study demonstrates that Habitat Radiomics features provide significant advantages over traditional Radiomics in predicting immunotherapy response in ESCC patients. The combined model, integrating both feature sets, shows exceptional predictive performance, with promising clinical applications in personalized treatment strategies. Future research will explore the broader applicability of this model across different cancer types and its integration with additional biomarkers to further enhance prediction accuracy.

The online version contains supplementary material available at 10.1186/s12967-025-07522-y.

Figures