An integrative deep learning–habitat radiomics model based on longitudinal CT for predicting treatment response to neoadjuvant immunochemotherapy in esophageal squamous cell carcinoma

Background Esophageal squamous cell carcinoma (ESCC) remains a major global health burden, with limited response rates to neoadjuvant immunochemotherapy (NICT). Accurate preoperative prediction of pathological complete response (pCR) may enable optimized treatment decisions. This study aimed to develop and validate an end-to-end deep learning–habitat fusion model integrating longitudinal CT images to predict pCR. Methods A total of 529 ESCC patients from two centers who received NICT followed by surgery were retrospectively analyzed. Longitudinal contrast-enhanced CT images were processed for tumor segmentation and habitat clustering using voxel-wise radiomic features. A transformer-based model (Habitat-Former) was designed to fuse temporal CT and habitat information for pCR prediction. Model performance was evaluated using area under the ROC curve (AUC). Prognostic value was assessed with Kaplan–Meier analyses. scRNA-seq was performed on 11 post-treatment tumor samples to explore immune-cell subpopulations and transcriptomic patterns potentially associated with pathological response. Results Habitat-Former achieved numerically higher performance across all cohorts, with AUCs of 0.917, 0.831, and 0.816 in the training cohort, internal test cohort, and external test cohort, respectively. In exploratory survival analyses among patients with complete follow-up, model-defined groups differed in PFS and OS in the training cohort ( P = 0.0076 and P = 0.037) but only in PFS in the internal test cohort ( P = 0.017). Exploratory scRNA-seq analysis of 11 post-treatment tumor samples identified candidate associations between model-derived imaging scores and myeloid-cell composition. Conclusions The proposed Habitat-Former model showed numerically favorable discrimination for pCR prediction in ESCC patients undergoing NICT and provided exploratory molecular correlates of treatment response in the tumor immune microenvironment. This model may serve as an adjunctive tool for preoperative response assessment and the development of imaging-driven biomarkers for immunotherapy response.

Authors

Institutions

Publication Details

Journal
Frontiers in Immunology
Published
2026-11-18
DOI
https://doi.org/10.3389/fimmu.2026.1912513
Primary Topic
Radiomics and Machine Learning in Medical Imaging
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An integrative deep learning–habitat radiomics model based on longitudinal CT for predicting treatment response to neoadjuvant immunochemotherapy in esophageal squamous cell carcinoma

Peng Tang, Ping Yuan, 姜宏景, Qingsong Pang et al.
7 links
Frontiers in Immunology
Radiomics and Machine Learning in Medical Imaging
article

An integrative deep learning–habitat radiomics model based on longitudinal CT for predicting treatment response to neoadjuvant immunochemotherapy in esophageal squamous cell carcinoma

Peng Tang, Ping Yuan, 姜宏景, Qingsong Pang, Wencheng Zhang, Xinyang Hu, Zhenhao Huang, Xi Chen, Tian Zhang
article en

Abstract

Background Esophageal squamous cell carcinoma (ESCC) remains a major global health burden, with limited response rates to neoadjuvant immunochemotherapy (NICT). Accurate preoperative prediction of pathological complete response (pCR) may enable optimized treatment decisions. This study aimed to develop and validate an end-to-end deep learning–habitat fusion model integrating longitudinal CT images to predict pCR. Methods A total of 529 ESCC patients from two centers who received NICT followed by surgery were retrospectively analyzed. Longitudinal contrast-enhanced CT images were processed for tumor segmentation and habitat clustering using voxel-wise radiomic features. A transformer-based model (Habitat-Former) was designed to fuse temporal CT and habitat information for pCR prediction. Model performance was evaluated using area under the ROC curve (AUC). Prognostic value was assessed with Kaplan–Meier analyses. scRNA-seq was performed on 11 post-treatment tumor samples to explore immune-cell subpopulations and transcriptomic patterns potentially associated with pathological response. Results Habitat-Former achieved numerically higher performance across all cohorts, with AUCs of 0.917, 0.831, and 0.816 in the training cohort, internal test cohort, and external test cohort, respectively. In exploratory survival analyses among patients with complete follow-up, model-defined groups differed in PFS and OS in the training cohort ( P = 0.0076 and P = 0.037) but only in PFS in the internal test cohort ( P = 0.017). Exploratory scRNA-seq analysis of 11 post-treatment tumor samples identified candidate associations between model-derived imaging scores and myeloid-cell composition. Conclusions The proposed Habitat-Former model showed numerically favorable discrimination for pCR prediction in ESCC patients undergoing NICT and provided exploratory molecular correlates of treatment response in the tumor immune microenvironment. This model may serve as an adjunctive tool for preoperative response assessment and the development of imaging-driven biomarkers for immunotherapy response.

Frontiers in ImmunologyVol. 17
Guangxi Medical University (CN), Tianjin Medical University Cancer Institute and Hospital (CN), First Affiliated Hospital of GuangXi Medical University (CN), First Affiliated Hospital of Zhengzhou University (CN), Tianjin Medical University (CN)
Good health and well-being
Radiomics and Machine Learning in Medical Imaging
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon