CTC-AttnNet: development and validation of an attention-based neural network for staging gastrointestinal malignancies
- 1Department of General Surgery, The Second Affiliated Hospital of Anhui Medical University, Hefei, Anhui China
- 2School of Computer Science and Technology, Soochow University, Suzhou, Jiangsu China
- 3The School of Public Health and Preventive Medicine, Monash University, Melbourne, VIC Australia
Abstract
Objective: To develop and validate a noninvasive model for classifying T stage in gastrointestinal malignancies using circulating tumor cells (CTCs) and serum tumor markers and to compare the predictive value of CTCs with that of conventional tumor markers.
Methods: This retrospective study included 328 patients with gastrointestinal malignancies from two hospitals in the same region. The internal modelling cohort comprised 214 patients (T1 + T2, n = 49; T3 + T4, n = 165), and the independent validation cohort comprised 114 patients (T1 + T2, n = 55; T3 + T4, n = 59). Univariable and multivariable logistic regression analyses were used to identify stage-related factors associated with CTCs. Three-dimensional-matched feature sets were constructed to minimize bias from unequal feature numbers. Repeated stratified fivefold cross-validation, the foldwise synthetic minority oversampling technique, and weighted cross-entropy loss were used to address severe class imbalance. Model performance was evaluated using AUC, balanced accuracy, F1 score, and calibration metrics rather than overall accuracy. Logistic regression, decision tree, support vector machine, k-nearest neighbors, and the newly developed CTC-AttnNet were compared. K‒M and Cox analyses were performed as exploratory recurrence analyses.
Results: CTC count alone moderately discriminated the overall TNM stage (AUC = 0.742) but poorly discriminated the T stage (AUC = 0.595); the AUCs of all four conventional tumor markers were < 0.60. Advanced overall TNM stage was independently associated with CTC positivity (OR = 1.844, 95% CI: 1.255–2.710; P = 0.002). Under dimension-matched conditions, the CTC-containing feature set achieved an external validation AUC of 0.76, outperforming the dimension-reduced conventional marker combination (AUC = 0.65). CTC-AttnNet achieved an external validation AUC of 0.82, a balanced accuracy of 0.78, and an F1 score of 0.76. Ablation analysis revealed that the attention module improved the recognition of minority class T1 + T2 cases. CTC positivity was also independently associated with recurrence (HR = 5.263, 95% CI 1.637–16.927; P = 0.005).
Conclusion: CTCs were significantly associated with gastrointestinal tumor stage and postoperative recurrence. Multivariable CTC-based models discriminated T stage better than models based on four conventional serum tumor markers did, although the CTC count alone had limited discriminatory ability. CTC-AttnNet represents an exploratory noninvasive stratification tool for this retrospective regional cohort and provides a methodological framework for developing liquid biopsy prediction models.
Supplementary Information: The online version contains supplementary material available at https://doi.org/10.1186/s12957-026-04530-0.
Related
- Development and external validation of risk prediction models for gastrointestinal bleeding after aortic dissection surgery.
- Development and Internal Validation of a Risk Prediction Model for Clozapine-Induced Gastrointestinal Hypomotility: A Single-Center Retrospective Cohort Study.
- Development and Initial Validation of the Multi-Dimensional Attention Rating Scale in Highly Educated Adults
- Advances in immunotherapy for gastrointestinal malignancies
- Second Primary Malignancies in Gastrointestinal Cancer Survivors: Epidemiology, Risk Factors, and Future Prediction Models
BibTeX
@article{li2026ctcattnnet,
title = {{CTC-AttnNet: development and validation of an attention-based neural network for staging gastrointestinal malignancies}},
author = {Liang Li and Chudi Sun and Shuiri Wang and Hao Qiang and Zhou Liu and Haojie Wang and Zhengtang Qi and Lin Zhang and Zhining Liu},
journal = {World Journal of Surgical Oncology},
year = {2026},
month = aug,
doi = {10.1186/s12957-026-04530-0},
url = {https://doi.org/10.1186/s12957-026-04530-0},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - CTC-AttnNet: development and validation of an attention-based neural network for staging gastrointestinal malignancies
AU - Liang Li
AU - Chudi Sun
AU - Shuiri Wang
AU - Hao Qiang
AU - Zhou Liu
AU - Haojie Wang
AU - Zhengtang Qi
AU - Lin Zhang
AU - Zhining Liu
PY - 2026
DA - 2026/08/01
JO - World Journal of Surgical Oncology
DO - 10.1186/s12957-026-04530-0
UR - https://doi.org/10.1186/s12957-026-04530-0
AB - Objective: To develop and validate a noninvasive model for classifying T stage in gastrointestinal malignancies using circulating tumor cells (CTCs) and serum tumor markers and to compare the predictive value of CTCs with that of conventional tumor markers. Methods: This retrospective study included 328 patients with gastrointestinal malignancies from two hospitals in the same region. The internal modelling cohort comprised 214 patients (T1 + T2, n = 49; T3 + T4, n = 165), and the independent validation cohort comprised 114 patients (T1 + T2, n = 55; T3 + T4, n = 59). Univariable and multivariable logistic regression analyses were used to identify stage-related factors associated with CTCs. Three-dimensional-matched feature sets were constructed to minimize bias from unequal feature numbers. Repeated stratified fivefold cross-validation, the foldwise synthetic minority oversampling technique, and weighted cross-entropy loss were used to address severe class imbalance. Model performance was evaluated using AUC, balanced accuracy, F1 score, and calibration metrics rather than overall accuracy. Logistic regression, decision tree, support vector machine, k-nearest neighbors, and the newly developed CTC-AttnNet were compared. K‒M and Cox analyses were performed as exploratory recurrence analyses. Results: CTC count alone moderately discriminated the overall TNM stage (AUC = 0.742) but poorly discriminated the T stage (AUC = 0.595); the AUCs of all four conventional tumor markers were < 0.60. Advanced overall TNM stage was independently associated with CTC positivity (OR = 1.844, 95% CI: 1.255–2.710; P = 0.002). Under dimension-matched conditions, the CTC-containing feature set achieved an external validation AUC of 0.76, outperforming the dimension-reduced conventional marker combination (AUC = 0.65). CTC-AttnNet achieved an external validation AUC of 0.82, a balanced accuracy of 0.78, and an F1 score of 0.76. Ablation analysis revealed that the attention module improved the recognition of minority class T1 + T2 cases. CTC positivity was also independently associated with recurrence (HR = 5.263, 95% CI 1.637–16.927; P = 0.005). Conclusion: CTCs were significantly associated with gastrointestinal tumor stage and postoperative recurrence. Multivariable CTC-based models discriminated T stage better than models based on four conventional serum tumor markers did, although the CTC count alone had limited discriminatory ability. CTC-AttnNet represents an exploratory noninvasive stratification tool for this retrospective regional cohort and provides a methodological framework for developing liquid biopsy prediction models. Supplementary Information: The online version contains supplementary material available at https://doi.org/10.1186/s12957-026-04530-0.
SN - 1477-7819
ER -