<p>Accurate segmentation in medical image analysis is essential for disease diagnosis and treatment planning, helping physicians precisely identify and locate lesions. The Segment Anything Model (SAM) is the first foundational model for general image segmentation, has shown impressive performance in various natural image segmentation tasks. However, medical image segmentation (MIS) is more challenging due to intricate anatomical structures, uncertain boundaries, and overall complexity of medical images. MedSAM’s introduction provides a general segmentation model for the medical image field. Nevertheless, general segmentation methods often struggle to effectively capture details and overall structures across different scales simultaneously and lack reliability assessments of predictions. To address these issues, we propose an improved MedSAM model. This model enhances perception of targets of various sizes using a multi-scale feature fusion module. It also employs Monte Carlo Dropout to estimate uncertainty, thereby increasing prediction reliability. Multi-scale feature fusion integrates information from various scales, optimizing detail capture and global structure understanding, while Monte Carlo Dropout assesses prediction uncertainty through multiple sampling, offering reliable assistance for clinical decision-making. Experimental results demonstrate that our method significantly improves accuracy and robustness across multiple medical image segmentation tasks.</p>