<p>In the current field of medical image segmentation research, numerous Transformer-based segmentation models have emerged. However, these models often suffer from limitations in multi-scale feature extraction and struggle to capture local detail features and contextual information, thereby constraining their segmentation performance. This paper introduces a novel model for medical image segmentation, called NVS-Former, which comprises both an encoder and a decoder. The key innovation of NVS-Former lies in its redesigned core module during the encoding phase, which not only enhances feature extraction capabilities but also improves the capture of local detail information. Additionally, the decoder structure has been reengineered to further optimize the model’s class prediction abilities. NVS-Former has demonstrated superior performance in tasks involving multi-organ, pulmonary detail, and cell segmentation. In various comparative experiments, it consistently outperformed state-of-the-art methods, highlighting its efficiency and stability in medical image segmentation.</p>