Enhancing Medical Image Segmentation with Adaptive Prompting with Foundation Model in Semi-Supervised Learning

Semi-supervised learning has emerged as a powerful solution for addressing the scarcity of labeled data in medical image segmentation. However, conventional methods often struggle with instability and inconsistency when handling unlabeled data, resulting in unreliable segmentation outcomes. To tackle these challenges, we introduce Adaptive Prompting with Segment Anything Model (ADP-SAM), a novel framework that integrates adaptive prompting and cross-teaching mechanisms. Adaptive prompting employs learnable prompts to guide the model's focus toward critical features, enhancing its ability to extract meaningful information from both labeled and unlabeled samples. Meanwhile, the cross-teaching mechanism utilizes two models that iteratively refine each other's outputs as pseudo-labels, reinforcing consistency and improving segmentation accuracy. Experimental evaluations on medical image datasets demonstrate that ADP-SAM consistently surpasses existing semi-supervised techniques, delivering more accurate and stable results. These findings highlight ADPSAM's potential to enable robust segmentation, particularly in scenarios where labeled data is limited.