Aim: This study aimed to explore the efficacy of MRI-based radiomics models, employing various machine learning techniques, in the preoperative prediction of the digital subtraction angiography (DSA) classification of venous malformations (VMs).
Materials and methods: In this retrospective study, 160 VM lesions from 153 children were categorized into a training set (n=128) and a testing set (n=32). Radiomic features were extracted from preoperative MRI scans. Feature selection was executed using the intraclass correlation coefficient test, z-scores, the K-best method, and the least absolute shrinkage and selection operator. Diverse MRI sequences and machine learning methods underpinned the development of the radiomics models. The models' efficacy was evaluated using receiver operating characteristic curves and the area under the curve (AUC).
Results: Out of 4528 radiomic features derived from CET1 and T2 images, 9 features were significantly associated with DSA classification differentiation. The most effective model for predicting VMs' DSA classification incorporated these 9 features and employed a random forest classifier. This model achieved an AUC of 0.917 in the training set and an excellent discrimination AUC of 0.891 in the testing set.
Conclusion: The random forest model, utilizing CET1 and T2 sequences, exhibited outstanding predictive performance in the preoperative distinction of VMs' DSA classification.
Copyright © 2025. Published by Elsevier Ltd.