PromptSeg: An End-to-End Universal Medical Image Segmentation Method via Visual Prompts.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models struggle with generalization in medical image segmentation due to diverse tasks and data variations.
• Task-specific models exhibit high uncertainty when applied to unseen data distributions, limiting their universal applicability.

Purpose of the Study:

• To introduce PromptSeg, a unified Transformer-based framework for universal 2D medical image segmentation.
• To leverage visual prompts and information theory to improve model generalization and reduce task uncertainty.

Main Methods:

• PromptSeg formulates segmentation as conditional entropy minimization, using visual prompts as side information.
• The framework is guided by the information bottleneck principle to filter noise and learn contextual representations.
• It requires minimal annotated prompt pairs for new tasks, enabling segmentation without retraining.

Main Results:

• PromptSeg demonstrates superior performance compared to state-of-the-art methods on various medical image segmentation tasks.
• The framework exhibits strong multi-modality generalization capabilities across different datasets (CT and MRI).
• PromptSeg effectively extracts task-specific semantics from visual prompts for accurate segmentation of query images.

Conclusions:

• PromptSeg offers a novel, unified approach to universal 2D medical image segmentation, overcoming generalization challenges.
• The prompt-based strategy allows for efficient adaptation to new tasks and modalities without extensive retraining.
• This framework represents a significant advancement in developing robust and versatile AI tools for medical image analysis.