Back to 2025 Abstracts
PREDICTION OF NODAL STATUS ON PREOPERATIVE MRI FOR PATIENTS WITH INTRAHEPATIC CHOLANGIOCARCINOMA: AN APPARENT DIFFUSION COEFFICIENT-BASED MACHINE LEARNING MODEL
Flavio Milana
*, Fabio Procopio, Giuseppe Ferrillo, Angela Ammirabile, Simone Famularo, Michele Fiordaliso, Eleonora Calafiore, Michela A. Polidoro, Jacopo Galvanin, Guido Costa, marco francone, Cristiana Bonifacio, Guido Torzilli
Surgery, IRCCS Humanitas Research Hospital, Rozzano, Lombardia, Italy
Background and Aims: Lymph node (N) status is a critical predictor of survival in patients with intrahepatic cholangiocarcinoma (iCCA). The benefit of lymphadenectomy remains debated when preoperative imaging shows no positive lymph nodes, given the risk of postoperative complications. This study aims to develop a machine-learning model to predict N status using preoperative MRI, potentially guiding surgical strategy.
Method: MRI data from a single-center cohort of iCCA surgical patients (2011-2023) were retrospectively analyzed. Two radiologists, blinded to surgical and pathological details, independently assessed lymph nodes ? 5mm in hepatic locoregional nodal stations. Lymph node diameter and Apparent Diffusion Coefficient (ADC) values were extracted from MRI scans obtained within two months before surgery. Inter-rater agreement was confirmed via Bland-Altman plots, allowing the calculation of ADC and diameter mean values among the two sets of observations. A Support Vector Machine (SVM) model with radial kernel was developed to classify N status using MRI along with clinical data extracted by logistic regression. Class weights to address class imbalance and leave-one-out cross-validation were performed. The primary endpoint was the model area under the curve (AUC), evaluated by ROC analysis. A 1000-fold bootstrap analysis provided internal validation.
Results: From 43 patients meeting the study criteria, 47 lymph nodes were matched to pathological reports, with 13 (27.7%) being metastatic. ADC values differed significantly between N0 and N1 groups (1177 vs 1001 mm2/s, p = 0.03), whereas diameter did not (14.2 vs 20.0mm, p = 0.18). Multifocal disease was the only clinical variable selected by multivariate logistic regression (OR 9.37, 95%CI: 1.69-74.8, p = 0.01). The SVM model, incorporating lymph node diameter, ADC values, and multifocal disease status, achieved an AUC of 0.91 (95%CI: 0.83-0.99) with 82.8% sensitivity, 92.3% specificity, 96.0% positive predictive value, and 70.6% negative predictive value
(Fig.1). The bootstrap internal validation resulted in an AUC of 0.87 (95%CI: 0.81-0.91)
(Fig.2).
Conclusion: Lymph node diameter alone does not reliably discriminate pathological status. Integrating diameter with ADC values, the proposed machine-learning model effectively predicts N status in the preoperative setting. This approach has the potential to refine treatment planning by reserving lymphadenectomy for patients most likely to benefit.
ROC curve for the Support Vector Machine model with the respective confusion matrix.
ROC Curve of the bootstrap model.
Back to 2025 Abstracts
