Three-dimensional anatomical classification and machine learning prediction of the P6a-right hepatic vein
Haoming Li1, Chengli Liu2,3,4, Jingchen Zhang5
1Department of Hepatobiliary Surgery, Air Force Medical University, Air Force Medical Center, Beijing, 100142, China.
Background:
The spatial relationship between the ventral branch of segment VI portal vein (P6a) and the right hepatic vein (RHV) significantly influences the anatomical complexity of the right hepatic fissure and the surgical outcomes of laparoscopic right anterior sectionectomy (LRAS). The ventral-P6a variant is often associated with higher surgical risks, yet preoperative prediction remains challenging.
Methods:
This retrospective study included 100 patients who underwent triple-phase enhanced CT of the upper abdomen. Based on three-dimensional reconstructions, P6a was classified as dorsal or ventral relative to the RHV according to its anatomical relationship. Five machine learning models-logistic regression, random forest, K-nearest neighbors, support vector machine, and XGBoost-were employed to predict P6a classification using anatomical angular and linear measurements. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and the F1-score, and the SHAP method was applied for feature interpretation.
Results:
Among the 100 patients, 59% were classified as dorsal-P6a and 41% as ventral-P6a. The ventral-P6a group exhibited a significantly smaller angle between the anterior and posterior portal branches (82.27° vs. 98.02°, P = 0.003), a higher proportion with a distance of less than 5 mm from the P6a origin to the posterior branch origin (37% vs. 10%, P = 0.001), and a larger RHV-inferior vena cava angle (52.23° vs. 43.97°, P = 0.013). The logistic regression model demonstrated the best performance in the test set (F1-score = 0.741) with minimal overfitting. The most important predictive features included the angle between the anterior and posterior portal branches, the angle of the right hepatic fissure plane, and the diameter of P6a. SHAP analysis validated the interpretability of the model.
Conclusion:
A machine learning model based on three-dimensional imaging anatomical features can effectively predict the spatial classification of P6a relative to the RHV. The interpretable logistic regression model identifies key anatomical risk factors, providing a basis for preoperative risk assessment and surgical planning in LRAS, particularly for optimizing surgical strategies in high-risk patients with the ventral-P6a variant.
