Reliable segmentation of arterial calcifications in CT angiography (CTA) is essential for stroke risk assessment and disease monitoring. Traditional threshold-based approaches, though simple and straightforward, often struggle with noise, contrast agent interference, and variability in plaque burden—leading to inconsistent or conservative segmentation results. In this work, we propose a semi-supervised learning framework based on the Mean Teacher paradigm to refine and improve segmentation quality derived from threshold-based pseudo-labels. We adapt the perturbation strategy for the teacher model to directly target the limitations observed in the threshold-based method, such as noise sensitivity and contrast-induced misclassifications. By introducing perturbations during the teacher’s inference, the model learns to counteract challenging conditions, guiding the student away from overfitting the noisy or inaccurate pseudo-labels. This approach encourages the network to unlearn artifacts from the threshold-based initialization and improves generalization to complex and variable imaging scenarios.