---
title: "IHU Strasbourg's S4M Beats SAM by +3.42 mIoU With Smarter 4-Point Prompts"
subtitle: "A structured 4-point prompting scheme with role-specific embeddings pushes SAM's medical segmentation accuracy up by 3.42 mIoU while cutting annotation time."
url: "https://alphasignal.ai/news/ihu-strasbourg-s-s4m-beats-sam-by-3-42-miou-with-smarter-4-point-prompts"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-13T00:00:00.000Z"
updated: "2026-08-24T02:10:21.724Z"
type: "Paper"
topics:
  - "Image"
  - "Open Source"
subtopics:
  - "Segmentation"
  - "Vision Language"
upvotes: 2456
source_url: "https://arxiv.org/abs/2503.05534"
comments: 0
truncated: true
---

# IHU Strasbourg's S4M Beats SAM by +3.42 mIoU With Smarter 4-Point Prompts

> A structured 4-point prompting scheme with role-specific embeddings pushes SAM's medical segmentation accuracy up by 3.42 mIoU while cutting annotation time.

Segment Anything (SAM) was supposed to end the annotation grind in medical imaging. Instead, clinicians clicking on ultrasound scans or surgical footage often find themselves stuck in refinement loops, because a single point on an organ with fuzzy edges tells the model almost nothing about where the boundary actually lies. A team from CAMMA at IHU Strasbourg argues the fix is a smarter prompt rather than a bigger model, and they have built [S4M](https://github.com/CAMMA-public/S4M) to prove it.

The core idea is disarmingly simple. Instead of clicking anywhere on the object, describe its shape with exactly four points that carry specific roles. The paper proposes a structured prompting strategy using four points as a compact instance-level shape description, and studies two variants: extreme points, and major/minor axis endpoints inspired by ultrasound measurement practice. The second variant matters because it already mirrors how sonographers work when they drop calipers on a scan.

## Why plain SAM stumbles in the clinic

SAM's point prompts are ambiguous in medical imaging because overlapping anatomy and blurred boundaries force cycles of manual refinement to reach a precise mask. You can throw more clicks at the problem, but each one is treated as an isolated hint. SAM cannot fully exploit structured prompts because it treats all points identically and lacks geometry-aware reasoning.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [S4M: 4-points to Segment Anything (arXiv)](https://arxiv.org/abs/2503.05534)
- [CAMMA-public/S4M on GitHub](https://github.com/CAMMA-public/S4M)
- [S4M: Segment Anything with 4 Extreme Points (v1)](https://arxiv.org/html/2503.05534v1)