Semi-Supervised medical image segmentation aims to improve model performance with minimal expert annotations, yet it faces challenges in maintaining consistent and high-quality learning. Excessive perturbations can distort the model’s predictions and disrupt the decision boundaries, particularly in regions with uncertain predictions. In this paper, we introduce Confidence-Aware Adaptive Displacement (CAD), a framework that selectively identifies and replaces the largest low-confidence regions with high-confidence patches. CAD dynamically adjusts both the maximum replacement size and the confidence threshold during training, progressively refining segmentation quality while avoiding overwhelming the learning process. Experimental results on public medical datasets demonstrate that CAD significantly enhances segmentation performance, achieving new state-of-the-art accuracy in this field.