Enhancing Cerebrovascular Segmentation With Explainable Deep Semi‐Supervised Learning: A Mean Teacher and Virtual Adversarial Training Approach

ABSTRACT Cerebral vascular segmentation is crucial for diagnosing and treating stroke. Although advances in deep learning have significantly enhanced segmentation, supervised models still rely on large, annotated data, coupled with their black‐box nature, presenting major challenges in medical settings. This study aims to develop an explainable deep semi‐supervised learning framework that incorporates Virtual Adversarial Training (VAT) and the Mean Teacher (MT) model. This approach leverages both labeled and unlabeled datasets to boost segmentation performance while reducing dependency on extensive manual annotations. In addition, techniques from Explainable AI (XAI), including Gradient‐weighted Class Activation Mapping (Grad‐CAM) and Monte Carlo (MC) Dropout, are incorporated to visualize feature regions and quantify uncertainty, enhancing model interpretability. Experiments reveal that the proposed model outperforms supervised U‐Net, the original MT, and the Uncertainty‐Aware Mean Teacher (UA_MT) model, achieving a Dice Similarity Coefficient (DSC) of 0.776 with superior generalization on testing data. Grad‐CAM visualizations confirm that the model correctly focuses on critical vascular structures, while the uncertainty maps identify areas with potential misclassification, guiding refinement and aiding clinical interpretation.