The unexpected pattern
Smaller colorectal tumors were associated with more metastasis, meaning cancer spread beyond the original tumor, after the comparison was statistically standardized. Using 5 cm as the dividing line, the crude overall metastasis rate was 57.8% in the smaller-tumor group and 56.2% in the larger-tumor group, a difference that was not statistically significant (P = 0.125). After standardization, a statistical adjustment intended to make the groups more comparable, the rates were 59.9% and 53.6%, respectively (P < 0.001). The result is an association in the study data, not evidence that a small tumor causes cancer to spread.
Size also did not behave like a simple rule in which every increase carried the same meaning. A flexible survival analysis found both an overall association between tumor length and overall survival and a nonlinear component, with P overall = 0.002 and P for non-linearity = 0.036. Within roughly 0 to 3 cm, the estimated mortality hazard rose as diameter fell, and the inverse pattern was stronger in male patients and those aged 50 or younger. The study's signal therefore varied across the size range rather than following one straight line.
A pattern concentrated in one subgroup
The survival contrast was clearest in a defined subgroup. Among male patients aged 50 or younger with T4 disease, those with tumors 3 cm or smaller had lower overall survival and disease-free survival than those with tumors larger than 5 cm. Both overall comparisons had P < 0.001. In patients without metastasis, the P value was 0.0026 for overall survival and below 0.001 for disease-free survival. Among patients with metastasis, the corresponding values were 0.92 and 0.053, so those differences did not reach statistical significance.
This was not a finding that small size carried the same prognostic meaning for every patient. The study examined selected surgical subgroups, and its evidence is observational. It therefore cannot establish that small tumors caused poorer survival or metastasis, or turn tumor size into a standalone clinical rule.
Prediction was useful, but less so elsewhere
The findings came from a retrospective study of 10,513 surgically treated colorectal cancer patients at FUSCC between 2008 and 2020. An independent cohort of 6,739 patients from Changhai Hospital was used for external validation. The researchers compared groups with standard statistical tests, used Kaplan-Meier and log-rank methods for survival comparisons, and used Cox regression to report hazard ratios with 95% confidence intervals.
For the prediction work, the primary target was lymph node metastasis. The team randomly split the cohort into stratified training and testing sets in an 8:2 ratio, then compared Decision Tree, Random Forest, ExtraTrees, XGBoost and CatBoost models. CatBoost produced the highest AUROC, 0.800. AUROC measures how well a model separates cases with and without the outcome. SHAP analysis, which shows how each input pushes a model's prediction up or down, gave smaller tumor diameter a positive contribution to predicted metastatic risk. That is a model signal, not a causal effect or an estimate of any one patient's absolute risk.
On the test data, CatBoost's AUROC was 0.800, with a 95% confidence interval of 0.782 to 0.818. Its Brier score was 0.179, with a 95% confidence interval of 0.170 to 0.186. The Brier score assesses how closely predicted probabilities match observed outcomes, so the two measures describe different aspects of model performance.
Performance fell in the independent Changhai cohort. AUROCs for the models ranged from 0.707 to 0.717, with Random Forest highest at 0.717, while Brier scores ranged from 0.204 to 0.207. The researchers described discrimination, calibration and net benefit as attenuated but still moderate. The weaker external performance shows that a model's results may change when it is applied to a different cohort.
Tumor length and survival
Survival modeling added a more familiar but less dramatic result. When tumor length was entered as a linear variable, greater length was associated with poorer overall survival. The unadjusted hazard ratio was 1.07, with a 95% confidence interval of 1.05 to 1.09 and P < 0.001. After adjustment, the hazard ratio was 1.04, with a 95% confidence interval of 1.01 to 1.06 and P = 0.004. A hazard ratio is a model-based comparison of the rate of an outcome over time; the authors described this effect as relatively small.
The broader message is narrower than a simple small-means-dangerous rule. Tumor size may carry information that a straightforward bigger-is-worse assumption misses, especially in the selected subgroup, but the evidence comes from retrospective surgical cohorts and the prediction findings weakened in external testing. The study does not show that small tumors cause metastasis or death, that size alone can determine prognosis, or that the pattern applies to every patient with colorectal cancer. Its practical caution is that a small tumor should not automatically be read as low risk in the subgroup where the adverse association was observed.
Paper data and sources
Original title: New insights into tumor size in colorectal cancer: smaller size, higher risk?
Authors: Ding R, Li T, Li C et al.
Journal/Repository: Frontiers in oncology
Status: Peer-reviewed
First online: 2026-10-19
DOI: Not available
Original paper · Full text