CT-based radiomics-clinical machine-learning model to predict completeness of cytoreduction in colorectal peritoneal metastases.
Completeness of cytoreduction (CC) remains the strongest prognostic determinant after cytoreductive surgery (CRS) ± hyperthermic intraperitoneal chemotherapy (HIPEC) for colorectal peritoneal metastases (CPM) yet accurate pre-operative prediction remains difficult. This study aimed to develop and validate a radiomic-clinical machine-learning model to predict cytoreduction completeness.
83 patients who underwent CRS ± HIPEC for CPM (2008-2025) were retrospectively analysed. Pre-operative contrast-enhanced CT scans were manually segmented in ITK-SNAP, and radiomic features were extracted using PyRadiomics. Clinical variables were modelled alone and in combination with radiomics features using a nested five-fold cross-validated machine-learning pipeline incorporating least absolute shrinkage and selection operator (LASSO) logistic regression, random forest (RF) and gradient-boosted classifiers (GBC) algorithms. The primary endpoint was incomplete cytoreduction (iCC). Model discrimination (AUROC, AUPRC), Brier score and calibration were assessed.
iCC occurred in 17 of 83 patients (20.5%). Independent predictors of iCC were high radiological PCI (≥15), upper-abdominal disease, absence of pre-operative chemotherapy and normal CEA (≤5 ng/mL). The radiomic-clinical model achieved the best performance (AUROC 0.90, AUPRC 0.69, Brier 0.077, sensitivity 0.83, specificity 0.92), outperforming clinical-only (AUROC 0.82-0.86) and radiomic-only (AUROC 0.69-0.75) models. Key radiomic predictors of iCC-low sphericity, high maximum 2D diameter and high zone entropy-reflected morphological irregularity and heterogeneity of CPM. Integrated models demonstrated superior calibration indicating stable and reliable probability estimates.
A CT-based radiomic-clinical model accurately predicts CC pre-operatively. This exploratory proof-of-concept model supports multicentre external validation to enhance decision-making for CRS ± HIPEC in CPM.
83 patients who underwent CRS ± HIPEC for CPM (2008-2025) were retrospectively analysed. Pre-operative contrast-enhanced CT scans were manually segmented in ITK-SNAP, and radiomic features were extracted using PyRadiomics. Clinical variables were modelled alone and in combination with radiomics features using a nested five-fold cross-validated machine-learning pipeline incorporating least absolute shrinkage and selection operator (LASSO) logistic regression, random forest (RF) and gradient-boosted classifiers (GBC) algorithms. The primary endpoint was incomplete cytoreduction (iCC). Model discrimination (AUROC, AUPRC), Brier score and calibration were assessed.
iCC occurred in 17 of 83 patients (20.5%). Independent predictors of iCC were high radiological PCI (≥15), upper-abdominal disease, absence of pre-operative chemotherapy and normal CEA (≤5 ng/mL). The radiomic-clinical model achieved the best performance (AUROC 0.90, AUPRC 0.69, Brier 0.077, sensitivity 0.83, specificity 0.92), outperforming clinical-only (AUROC 0.82-0.86) and radiomic-only (AUROC 0.69-0.75) models. Key radiomic predictors of iCC-low sphericity, high maximum 2D diameter and high zone entropy-reflected morphological irregularity and heterogeneity of CPM. Integrated models demonstrated superior calibration indicating stable and reliable probability estimates.
A CT-based radiomic-clinical model accurately predicts CC pre-operatively. This exploratory proof-of-concept model supports multicentre external validation to enhance decision-making for CRS ± HIPEC in CPM.
Authors
Pau Pau, Eglinton Eglinton, Wang Wang, Mehri-Kakavand Mehri-Kakavand, Fischer Fischer
View on Pubmed