Related Experiment Video
Updated: May 29, 2026
Establishing a Physiologic Human Vascularized Micro-Tumor Model for Cancer Research
Published on: September 15, 2023
Interpretable and non-linear machine learning models for quantitative ICG perfusion assessment in colorectal cancer
Paulina Daniluk-Marsy1, Johannes Castelein2, Karol Polom3
1Department of Surgical Oncology, Transplant Surgery and General Surgery, Faculty of Medicine, Medical University of Gdańsk, Poland.
Background:
Anastomotic leakage (AL) remains a major complication after colorectal cancer surgery (CCS). Intraoperative indocyanine green (ICG) fluorescence angiography is widely used to assess bowel perfusion, but interpretation is largely subjective. Quantitative analysis of fluorescence kinetics combined with machine learning (ML) may enable more objective, physiology-based risk stratification.
Materials And Methods:
In this single-centre retrospective feasibility study, 81 patients undergoing colorectal cancer resection with ICG perfusion assessment were analysed. Intraoperative fluorescence videos were post-processed to extract four quantitative perfusion metrics. Eight clinical variables with plausible relevance to microvascular health were included. Intraoperative hypoperfusion (n = 12, 15%) was defined by delayed fluorescence propagation judged by the operating surgeon. Logistic regression with elastic-net regularisation and three non-linear algorithms (random forest, XGBoost, support vector machine) were trained using perfusion metrics-only, clinical variables-only, and combined feature sets. Performance was evaluated with repeated stratified 5-fold cross-validation and bootstrap confidence intervals.
Results:
Using perfusion-only features, elastic-net logistic regression achieved a ROC-AUC of 0.76 (95% CI 0.62-0.88). Clinical-only models reached a ROC-AUC of 0.77 (95% CI 0.59-0.93) with random forest. The combined model (perfusion + clinical) yielded the highest discrimination (ROC-AUC 0.81, 95% CI 0.63-0.94) with good calibration (Brier score 0.14). SHAP analysis identified increased latency and lower plateau intensity ratio (PIR) as the strongest predictors of hypoperfusion.
Conclusion:
Quantitative ICG perfusion indices, particularly latency and PIR, combined with interpretable ML models, can predict surgeon-assessed intraoperative hypoperfusion during colorectal cancer surgery with good discrimination. These findings support further development of explainable, perfusion-guided decision-support tools to reduce anastomotic risk.
