ProSona: Prompt-Guided Personalization for Multi-Expert Medical Image Segmentation

The introduction of ProSona marks a significant advancement in automated medical image segmentation, particularly for challenging tasks such as lung nodule delineation, where expert opinions often vary. This two-stage framework leverages a continuous latent space of annotation styles, allowing for personalized segmentation guided by natural language prompts. The results demonstrate a 17% reduction in Generalized Energy Distance and an improvement of over one point in mean Dice scores compared to the previous model, DPersona. These enhancements underscore the effectiveness of ProSona in providing accurate and interpretable segmentation, which is crucial for clinical decision-making. As ProSona is made available online, it opens new avenues for integrating AI in medical imaging, potentially transforming how radiologists approach image analysis.