EviPrompt: A Training-Free Evidential Prompt Generation Method for Adapting Segment Anything Model in Medical Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is vital for clinical applications.
• The Segment Anything Model (SAM) shows promise but faces challenges in medical imaging due to prompt engineering and domain gaps.
• Existing methods often require extensive labeling and computational resources.

Purpose of the Study:

• To introduce EviPrompt, a novel, training-free prompt generation method for medical image segmentation.
• To overcome the limitations of SAM in medical imaging, specifically the need for expert prompts and the domain gap.
• To reduce the reliance on extensive labeling and computational resources.

Main Methods:

• EviPrompt automatically generates prompts using a single reference image-annotation pair.
• It employs evidential learning for improved prompt reliability.
• Committee voting and inference-guided in-context learning are used to address the domain gap, prioritizing human prior knowledge.

Main Results:

• EviPrompt demonstrates efficacy across a diverse range of medical imaging tasks and modalities.
• The method significantly reduces the need for expert labor and computational resources.
• It provides an efficient and robust solution for medical image segmentation.

Conclusions:

• EviPrompt offers an effective and efficient solution for medical image segmentation.
• The training-free approach makes advanced segmentation accessible with minimal resources.
• This method has the potential to significantly impact clinical applications of medical imaging.