S4M: 4-points to segment anything.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• The Segment Anything Model (SAM) offers potential for medical image segmentation.
• Current SAM prompting methods struggle with ambiguous anatomy and blurred boundaries, requiring extensive manual refinement.
• There is a need for improved prompting strategies to enhance SAM's utility in medical applications.

Purpose of the Study:

• To develop a structured prompting strategy for medical image segmentation using minimal annotation cost.
• To augment the Segment Anything Model (SAM) to interpret 4-point prompts as relational cues, not isolated clicks.
• To introduce S4M (4-points to Segment Anything) for geometry-aware reasoning in medical segmentation.

Main Methods:

• Proposed a structured prompting strategy using 4-points (extreme points and major/minor axis endpoints) for instance-level shape description.
• Introduced S4M, which enhances SAM by incorporating role-specific embeddings for relational point interpretation.
• Utilized an auxiliary "Canvas" pretext task to strengthen prompt representation learning and foster geometry-aware reasoning.

Main Results:

• S4M achieved a +3.42 mIoU improvement over a strong SAM baseline across eight diverse medical imaging datasets.
• Annotation studies confirmed that major/minor axis prompts facilitate faster and more practical annotation by clinicians.
• The proposed method demonstrates significant gains in segmentation performance with minimal annotation effort.

Conclusions:

• S4M enhances segmentation performance and reduces annotation effort in medical imaging.
• The strategy aligns with existing clinical workflows, facilitating adoption and scalable dataset development.
• The S4M framework offers a promising approach for improving medical image segmentation accuracy and efficiency.