K-Prism: A Knowledge-Guided and Prompt Integrated Universal Medical Image Segmentation Model

K-Prism is a novel universal medical image segmentation model that integrates three knowledge paradigms: semantic priors from annotated datasets, in-context knowledge from few-shot examples, and interactive feedback from user inputs. It employs a dual-prompt representation with 1-D sparse prompts for segmentation tasks and 2-D dense prompts for attention guidance, utilizing a Mixture-of-Experts (MoE) decoder for dynamic routing. This architecture allows for flexible paradigm switching and joint training across various tasks, achieving state-of-the-art performance on 18 public datasets across multiple modalities, which is crucial for practitioners aiming for comprehensive segmentation solutions in clinical settings.