Diagnostic Accuracy StudyEditorially reviewed

An interpretable CT-based machine learning model for predicting recurrence risk in stage II colorectal cancer.

Ziqi Wu, Liya Gong, Jingwen Luo, Xiaobo Chen, Fan Yang, Junyan Wen, Yanyu Hao, Zhishan Wang, Ruozhen Gu, Yuqin Zhang, Hai Liao, Ge Wen

Insights into Imaging · July 31, 2025 · Vol 16 · Issue 1 · p. 162

PMID: 40745419PMCID: PMC12314294DOI: 10.1186/s13244-025-02009-2

Open original article

This is an evidence-grounded AI reading aid that has passed human editorial review. Items markedRadWeave readingare our interpretation rather than statements by the authors. It is not a substitute for reading the original publication, and RadWeave does not reproduce the source full text here.

Take-home message

A multicenter ANN-based combined model integrating CT radiomics and clinicoradiological features stratified stage II colorectal cancer patients into high- and low-risk groups for three-year disease-free survival, outperforming both radiomics-only and clinicoradiological-only models across test and external validation cohorts.

SHAP analysis identified Radscore, age, and rough outer edge of the intestine as the top three contributors to the combined model's predictions, followed by ctEMVI, perineural invasion, and intestinal obstruction or perforation status. The outer edge of the intestine was the single most important clinicoradiological predictor.

What the radiologist should know

Rough or irregular outer edge of the intestine on CT was identified as the leading clinicoradiological predictor of recurrence risk in this model, reflecting tumor aggressiveness and deeper tissue invasion. Readers should be aware this association was established within a predictive model context and does not yet constitute a validated standalone reporting standard.
Evidence grounded · 2 source references
Five clinicoradiological factors—age, perineural invasion, intestinal obstruction or perforation status, outer edge of the intestine, and ctEMVI—were identified as independent predictors of recurrence by multivariate logistic regression and incorporated into the clinicoradiological model. The clinicoradiological model alone showed modest and declining discrimination across cohorts.
Evidence grounded · 2 source references
Kaplan–Meier analysis confirmed statistically significant separation of high- and low-risk groups across all three cohorts, with the high-risk group showing substantially worse disease-free survival. The hazard ratio in the training cohort was notably larger than in the test cohort, suggesting some optimism in internal estimates.
Evidence grounded · 1 source reference

Reporting implications

This paper does not support a change to reporting.

No established routine reporting change is supported. Feature awareness: CT outer edge morphology and ctEMVI are model inputs with recurrence-risk signal, not yet validated as standalone reporting standards. Prospective confirmation and clinical actionability evidence are needed before routine reporting change.

Practice impact

RadWeave editorial assessment, not a statement by the authors.

Potentially useful

The combined ANN model shows promising discrimination for stage II CRC recurrence risk stratification, but remains investigational and unvalidated prospectively. No routine reporting or management change is established.

Evidence grounded · 3 source references

Caveats before applying this

The retrospective design and single-phase (portal venous) CT acquisition mean reported performance may not reflect prospective or multi-phase practice; arterial and noncontrast phases were not evaluated, limiting generalisability.
Evidence grounded · 1 source reference
Adding radiomics to clinicoradiological factors significantly improved NRI and IDI in the validation cohort, but the combined model's AUC advantage over the radiomics-only model did not reach statistical significance in the validation cohort (DeLong p = 0.053), warranting cautious interpretation of incremental gain.
Evidence grounded · 2 source references
The notably larger hazard ratio in the training cohort compared with the test cohort suggests some optimism in internal performance estimates, and external validation in independent prospective cohorts is needed before clinical adoption.
RadWeave reading
Evidence grounded · 1 source reference

Numbers worth remembering

Combined model — AUC: 0.811
Combined model — AUC: 0.846
High-risk vs low-risk — Hazard ratio: 7.364 (95% CI: 5.342–10.151)
Clinicoradiological model — AUC: 0.706
Clinicoradiological model — AUC: 0.608
Clinicoradiological model — AUC: 0.572
training cohort — Sample size: 442 patients

RadWeave bottom line

This multicenter ANN combining CT radiomics and clinicoradiological features shows promising discrimination for three-year recurrence risk in stage II colorectal cancer, but remains investigational; no change in routine CT reporting practice is established pending prospective validation.

RadWeave reading

Evidence grounded · 4 source references

Study in 20 seconds

Study type: Diagnostic Accuracy Study
Population: Stage II colorectal cancer patients from three hospitals undergoing curative resection, with three-year follow-up for disease-free survival
Modality: CT (portal venous phase)
Technique: CT radiomics combined with clinicoradiological features in an artificial neural network (ANN) model, with SHAP explainability analysis
Comparator / reference: Radiomics-only model and clinicoradiological-only model
Primary endpoint: Three-year disease-free survival risk stratification (high vs low risk) in stage II CRC