Segment anything in medical images

Jun Ma; Yuting He; Feifei Li; Lin Han; Chenyu You

doi:10.1038/s41467-024-44824-z

Home SearchPaper Details

Segment anything in medical images

Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, Bo Wang

Nature Portfolio (2024) • Volume 15, Issue 1, Pages 654-654

Method-ToolMethod-ToolPDF AvailableGrade Eligible

Overall Assessment

Strong Methodological Quality

Assessment created by PaperScorers Medical AI v0.1.0 on Dec 14, 2025

B+

81/100

Key Takeaways

•MedSAM fine-tunes SAM on 1.57M med image–mask pairs across 10 modalities.
•Outperforms SAM and often specialist U-Net/DeepLabV3+ on 86 internal and 60 external tasks.
•External generalisation strong, incl. unseen targets/modalities.
•Training scale improves performance; annotation time cut by ~82% in user study.
•Open code/model and dataset links; no COIs declared.

Conclusion

Robust, broadly generalisable segmentation foundation model with strong benchmarking and good transparency, albeit limited stats correction and no uncertainty reporting.

Quick Actions

Read Full Paper

Quality Dimensions

Bias & Integrity

COI handling, outcome switching, selective reporting

B+84

Methods Rigour

Protocol clarity, sampling, controls

B+82

External Validity

Generalisability, setting, population

A-86

Transparency & Reproducibility

Data/code/materials availability, preregistration

A-88

Significance & Novelty

Contribution vs prior art; context if known

A90

Statistical Validity

Models, assumptions, multiplicity, CI priority

C62

Integrity & Transparency

Integrity checks

P-hacking risk?

?

None suspected

Signs of selective hypothesis testing or analysis choices that could inflate false positive rates.

Outcome switching?

?

None suspected

Evidence that primary outcomes were changed during or after the study, potentially distorting results.

Conflict of interest?

?

None suspected

Authors have financial or other relationships that could bias research findings.

Data integrity issues?

?

None suspected

Concerns about the accuracy, completeness, or authenticity of the reported data.

Open science signals

Research data

?

Open access

Raw data used in the analyses is publicly available or accessible upon request.

Analysis code

?

Open access

Code or scripts used to analyze the data are shared for reproducibility.

Study materials

?

Open access

Protocols, questionnaires, and other materials are publicly available.

Premise

Primary Research Question

Can a promptable foundation model fine-tuned on medical images deliver universal segmentation outperforming SAM and specialist models?

Hypothesis

Fine-tuning SAM on a very large medical dataset yields superior, generalisable segmentation across modalities.

Hypothesis is Falsifiable

✓Yes

Tested via 86 internal and 60 external tasks with quantitative metrics.

PICO Framework

Population—Not specified

Method/benchmark, not a patient/intervention study.

Intervention—Not specified

No clinical intervention; algorithmic comparisons only.

Comparison—Not specified

Clinical PICO not applicable; comparisons are methodological.

Outcome—Not specified

Clinical outcomes not studied; segmentation metrics used.

Literature Positioning

Literature Review Balanced

Well covered

Positions against SAM and specialist models; cites surveys and SOTA.

Evidence: pp. 1–2

Research Gap Clearly Stated

Well covered

Need for universal, promptable medical segmentation foundation model.

Evidence: pp. 1–2

Stated Contribution Clear

Well covered

Introduce MedSAM; trained on 1.57M pairs; superior accuracy and robustness.

Evidence: Abstract p. 1

Study Provenance

Peer reviewed venueFunding disclosedNo conflicts declared5 affiliations listed

Authors & Affiliations

UHN/UofT/Vector Institute; Western Univ; NYU Tandon; Yale E.E.; UHN AI Hub.

Funding Statement

NSERC RGPIN-2020-06189, DGECR-2020-00294; CIFAR AI Chair programmes.

Conflicts of Interest

Authors declare no competing interests.

Peer Review

✓Nature Communications (peer-reviewed).

Evidence: p. 9

Methodological Assessment

Modeling / SimulationModel ReportingModeling

5 applicable checklist items.

Yes 4 · No 1

✓

Structure Transparent

Architecture, training hyperparameters, prompts, losses, and compute fully detailed.

Evidence: Methods: Network architecture & Training protocol, pp. 5–7

✗

Uncertainty Propagated

No predictive uncertainty quantification reported.

Evidence: Methods/Results

✓

Sensitivity Analyses (Global)

Scaling law (10k/100k/1M) and annotation study; broad ablations.

Evidence: Fig. 5a–b

✓

External Validation Performed

60-task external validation on unseen datasets/targets/modalities.

Evidence: Fig. 4a–c

✓

Code/Model Available

GitHub + Zenodo release of scripts and trained model.

Evidence: Code availability p. 5; Ref. 47

Modules without applicable checklist items (9)

These designs were fully marked as not applicable but remain available for reference.

Quick Actions

Read Full Paper

Study Overview

Study Design

Method Tool

Primary classification used to evaluate this research.

Population

Public medical imaging datasets across 10 modalities; >30 cancer types; 86 internal and 60 external tasks.

Brief summary of who or what was studied.

Sample Size

1,570,263

Number of participants included in the study analyses.

Study Arms

4 arms

Total number of intervention or comparison groups.

Pre-registration

Not registered

Indicates whether the protocol or trial was registered before data collection.

Primary Outcome

Primary outcome was pre-specified in the protocol or registry.

Effect Size

DSC/NSD performance: Nasopharynx cancer: DSC 87.8% (IQR 85.0–91.4), +52.3 vs SAM

Reviewer Notes

Foundation segmentation model development with internal/external validation; benchmarking vs SAM/U-Net/DeepLabV3+.

Publication Details

DOI

10.1038/s41467-024-44824-z

Published

January 22, 2024

Citations

1654 citations

FWCI

1028.43

External Resources

DOI Link OpenAlex

Disclaimer: This assessment is generated by AI and should not be the sole basis for clinical or research decisions. Always review the original paper and consult with domain experts.

Suggested Papers

D+

Long COVID and the Increased Risk of Food Insecurity Among Participants in Arizona CoVHORT: A matched cohort study

Chidera M. Ejike et al.•2025

B+

2023 AHA/ACC/ACCP/ASPC/NLA/PCNA Guideline for the Management of Patients With Chronic Coronary Disease: A Report of the American Heart Association/American College of Cardiology Joint Committee on Clinical Practice Guidelines

Salim S. Virani et al.•2023

No Report

Hyperfunctional C3 convertase leads to complement deposition on endothelial cells and contributes to atypical hemolytic uremic syndrome

Lubka T. Roumenina et al.•2009

Suggested Papers

From Our Blog

Regression to the Mean: The Illusion of Improvement

Why do sick people get better after taking a useless pill? It might just be regression to the mean. Don't be fooled by natural fluctuations.

Standard Deviation: Why Average is Boring

If you put your head in the oven and your feet in the freezer, on average you are comfortable. You are also dead.

HARKing: Rewriting History in Science

Hypothesizing After the Results are Known. It makes you look like a genius. It makes your science worthless.