Machine Learning-Based Prediction of Fibular Artery Septocutaneous Perforators.
Authors
Affiliations (4)
Affiliations (4)
- TUM University Hospital Klinikum Rechts der Isar, School of Medicine and Health, Technical University of Munich, Department of Oral and Maxillofacial Surgery, Bavaria, Germany, Munich.
- TUM University Hospital Klinikum rechts der Isar, School of Medicine and Health, Technical University of Munich, Institute of Diagnostic and Interventional Radiology, Bavaria, Germany, Munich.
- Technical University of Munich and TUM University Hospital, Chair for AI in Healthcare and Medicine, Bavaria, Germany, Munich.
- University Hospital Leipzig, Department of Oral and Maxillofacial Surgery, Saxony, Germany, Leipzig.
Abstract
Reliable identification of septocutaneous fibular artery perforators is important for free fibula flap harvest involving a skin paddle. While imaging modalities such as computed tomography angiography (CTA) provide high diagnostic accuracy for direct perforator visualization, they do not quantify segment-level probability of perforator presence based on global anatomical characteristics. In this retrospective CTA-based study, 246 patients (490 limbs) were analyzed. Each fibula was divided into 10 equal segments, yielding 4,900 limb-zone observations. Using limb-level anatomical and vascular parameters, machine learning models were developed to predict the presence of at least one perforator per segment. Data were split at the patient level (80% training, 20% testing). Model performance was evaluated using area under the receiver operating characteristic curve (ROC-AUC), area under the precision-recall curve (PR-AUC), and Brier score. Overall segment-level perforator prevalence was 15.4%, and 20.6% of limbs exhibited no septocutaneous perforator. Ridge-regularized logistic regression (primary model) achieved a ROC-AUC of 0.684, PR-AUC of 0.243, and Brier score of 0.1214. Random forest achieved a ROC-AUC of 0.705, PR-AUC of 0.239, and Brier score of 0.1216. XGBoost achieved a ROC-AUC of 0.710, PR-AUC of 0.245, and Brier score of 0.1213. The generalized additive model yielded a ROC-AUC of 0.683, PR-AUC of 0.231, and Brier score of 0.1223. All approaches reproduced the characteristic midfibular peak in perforator probability. Machine learning enables moderate segment-level discrimination of perforator presence based on global anatomical features. While not comparable to imaging for direct vessel localization, probabilistic modeling may complement imaging by providing structured, patient-specific estimation of regional perforator likelihood.