Voxel-level habitat radiomics for postoperative recurrence risk stratification in double-invasion gastric cancer

To develop and internally validate a voxel-level habitat radiomics framework for postoperative recurrence risk stratification in gastric cancer with pathologically confirmed double invasion, and to examine the biological plausibility of imaging-defined habitat patterns using representative immunohistochemistry and orthogonal TCGA-STAD transcriptomic analysis. In this multicenter retrospective study, we enrolled 327 patients with gastric adenocarcinoma and postoperative pathological evidence of concurrent vascular and perineural invasion from three institutions and divided them into training ( n = 229) and internal validation ( n = 98) cohorts. All patients underwent preoperative three-phase contrast-enhanced CT (arterial, portal venous, and delayed phases) within 4 weeks before surgery. We segmented tumors on portal venous-phase CT, then transferred the ROI to the arterial and delayed phases after rigid registration and visual quality checking. For voxel-level habitat partitioning, we used a training-derived global K-means model fitted to pooled tumor voxels from the training cohort; we fixed the learned centroids for validation. We selected habitat-specific radiomic features only from the training cohort and integrated them with clinicopathological variables to build recurrence-free survival models. We derived the upstream image-processing and feature-selection steps from the training cohort; an audit of the archived notebooks showed that Cox ridge penalties were selected using the internal validation subset, making its performance estimates conditional on that tuning; we did not apply ComBat harmonization in the primary reported models. We assessed performance using the C-index; inverse-probability-of-censoring weighted (IPCW) cumulative/dynamic time-dependent ROC analysis at 36 and 60 months; patient-level paired bootstrap AUC comparisons (1,500 resamples); calibration; decision-curve analysis; reclassification metrics; a 1,000-resample fixed-feature Cox refitting sensitivity audit of training-sample optimism; within-cohort fixed-score bootstrap confidence intervals; and center-stratified analyses. We used three-year RFS as the primary prediction horizon because follow-up was more mature at this time point, while retaining 5-year analyses as secondary exploratory estimates. We assessed biological plausibility using representative immunohistochemistry and TCGA-STAD transcriptomic ecosystem scores. Habitat decomposition identified four imaging-defined subregions interpreted as angiogenic-like, hypoxic-like, stromal-remodeling-like, and immune-associated habitats. The habitat radiomics model achieved C-index values of 0.733 and 0.728 in the training and internal validation cohorts, respectively. The combined model showed only marginally higher C-index values of 0.735 and 0.730, indicating that the habitat signature carried most of the discriminative information. At 3 years, IPCW AUCs were 0.803/0.874 for the habitat model, and 0.807/0.881 for the combined model (training/internal validation); paired combined-minus-habitat AUC differences were small, and their 95% bootstrap confidence intervals included zero. In contrast, the whole-tumor radiomics model showed limited validation performance. Calibration and decision-curve analyses suggested acceptable, provisional internal performance and potential decision utility within the studied cohort. TCGA-STAD analysis provided orthogonal biological support for angiogenic, stromal, and composite ecosystem programs as adverse prognostic features, with adjusted HRs of 1.30, 1.34, and 1.14, respectively. In gastric cancer with concurrent vascular and perineural invasion, three-phase CT habitat radiomics captured spatial prognostic information that whole-tumor radiomics did not represent as effectively. The habitat signature showed higher 3-year RFS discrimination in internal validation and accounted for nearly all of the combined model’s discriminative performance. These findings establish a focused imaging framework for refining postoperative risk stratification within a pathologically defined high-risk population and support future evaluation of habitat-informed surveillance strategies.

Authors

Institutions

Publication Details

Journal
Journal of Translational Medicine
Published
2026-11-18
DOI
https://doi.org/10.1186/s12967-026-09178-8
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

Voxel-level habitat radiomics for postoperative recurrence risk stratification in double-invasion gastric cancer

黄庆先, 马晋峰, Wenhua Cheng, Yifan Li et al.
4 links
Journal of Translational Medicine
Radiomics and Machine Learning in Medical Imaging
article

Voxel-level habitat radiomics for postoperative recurrence risk stratification in double-invasion gastric cancer

黄庆先, 马晋峰, Wenhua Cheng, Yifan Li, Gaopeng Li, Jianxin Zhang, Chengyan Zhang, Yan Liu
article en

Abstract

To develop and internally validate a voxel-level habitat radiomics framework for postoperative recurrence risk stratification in gastric cancer with pathologically confirmed double invasion, and to examine the biological plausibility of imaging-defined habitat patterns using representative immunohistochemistry and orthogonal TCGA-STAD transcriptomic analysis. In this multicenter retrospective study, we enrolled 327 patients with gastric adenocarcinoma and postoperative pathological evidence of concurrent vascular and perineural invasion from three institutions and divided them into training ( n = 229) and internal validation ( n = 98) cohorts. All patients underwent preoperative three-phase contrast-enhanced CT (arterial, portal venous, and delayed phases) within 4 weeks before surgery. We segmented tumors on portal venous-phase CT, then transferred the ROI to the arterial and delayed phases after rigid registration and visual quality checking. For voxel-level habitat partitioning, we used a training-derived global K-means model fitted to pooled tumor voxels from the training cohort; we fixed the learned centroids for validation. We selected habitat-specific radiomic features only from the training cohort and integrated them with clinicopathological variables to build recurrence-free survival models. We derived the upstream image-processing and feature-selection steps from the training cohort; an audit of the archived notebooks showed that Cox ridge penalties were selected using the internal validation subset, making its performance estimates conditional on that tuning; we did not apply ComBat harmonization in the primary reported models. We assessed performance using the C-index; inverse-probability-of-censoring weighted (IPCW) cumulative/dynamic time-dependent ROC analysis at 36 and 60 months; patient-level paired bootstrap AUC comparisons (1,500 resamples); calibration; decision-curve analysis; reclassification metrics; a 1,000-resample fixed-feature Cox refitting sensitivity audit of training-sample optimism; within-cohort fixed-score bootstrap confidence intervals; and center-stratified analyses. We used three-year RFS as the primary prediction horizon because follow-up was more mature at this time point, while retaining 5-year analyses as secondary exploratory estimates. We assessed biological plausibility using representative immunohistochemistry and TCGA-STAD transcriptomic ecosystem scores. Habitat decomposition identified four imaging-defined subregions interpreted as angiogenic-like, hypoxic-like, stromal-remodeling-like, and immune-associated habitats. The habitat radiomics model achieved C-index values of 0.733 and 0.728 in the training and internal validation cohorts, respectively. The combined model showed only marginally higher C-index values of 0.735 and 0.730, indicating that the habitat signature carried most of the discriminative information. At 3 years, IPCW AUCs were 0.803/0.874 for the habitat model, and 0.807/0.881 for the combined model (training/internal validation); paired combined-minus-habitat AUC differences were small, and their 95% bootstrap confidence intervals included zero. In contrast, the whole-tumor radiomics model showed limited validation performance. Calibration and decision-curve analyses suggested acceptable, provisional internal performance and potential decision utility within the studied cohort. TCGA-STAD analysis provided orthogonal biological support for angiogenic, stromal, and composite ecosystem programs as adverse prognostic features, with adjusted HRs of 1.30, 1.34, and 1.14, respectively. In gastric cancer with concurrent vascular and perineural invasion, three-phase CT habitat radiomics captured spatial prognostic information that whole-tumor radiomics did not represent as effectively. The habitat signature showed higher 3-year RFS discrimination in internal validation and accounted for nearly all of the combined model’s discriminative performance. These findings establish a focused imaging framework for refining postoperative risk stratification within a pathologically defined high-risk population and support future evaluation of habitat-informed surveillance strategies.

Journal of Translational Medicine
Shanxi Medical University (CN), Chinese Academy of Medical Sciences & Peking Union Medical College (CN), Shanxi Provincial Cancer Hospital (CN), Shanxi Academy of Medical Sciences (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