Abstract: Effective multi-scale feature representation is crucial for accurately segmenting lesions of varying sizes in medical images. Addressing the challenge of existing methods failing to fully exploit multiscale information for different targets, a multi-scale perception segmentation network is proposed. It first explores the perceptual ability of multiscale contextual information from both local and global perspectives, constructing multiscale encoders and decoders. The multi-scale encoder utilizes a local multi-scale self-attention mechanism and global fine-tuning to extract features at multiple scales, capturing information from different targets in the image. The multi-scale decoder, through upsampling, restores spatial resolution while preserving detailed information, leading to more precise segmentation results. To further enhance feature semantic representation, a semantic diffusion alignment module is introduced, achieving semantic alignment between low-level and high-level features to obtain more discriminative fused features. Experimental validation on multiple medical image datasets of different modalities demonstrates the outstanding performance of the proposed method, surpassing most current medical image segmentation methods and achieving more accurate and robust segmentation results.