The DOI matches its Crossref record; 51 of the 51 references it cites were found in a registry or an index; 13 of the 26 sources its claims cite were checked for support.
Segment anything in medical images
Open accessAbstract
Medical image segmentation is a critical component in clinical practice, facilitating accurate diagnosis, treatment planning, and disease monitoring. However, existing methods, often tailored to specific modalities or disease types, lack generalizability across the diverse spectrum of medical image segmentation tasks. Here we present MedSAM, a foundation model designed for bridging this gap by enabling universal medical image segmentation. The model is developed on a large-scale medical image dataset with 1,570,263 image-mask pairs, covering 10 imaging modalities and over 30 cancer types. We conduct a comprehensive evaluation on 86 internal validation tasks and 60 external validation tasks, demonstrating better accuracy and robustness than modality-wise specialist models. By delivering accurate and efficient segmentation across a wide spectrum of tasks, MedSAM holds significant potential to expedite the evolution of diagnostic tools and the personalization of treatment plans.
Overview
CoChat Summary
TL;DR
MedSAM is a promptable foundation model for universal medical image segmentation, developed by fine-tuning SAM on 1,570,263 medical image-mask pairs across 10 imaging modalities and over 30 cancer types. It outperformed SAM and often matched or surpassed modality-wise specialist models across 86 internal and 60 external validation tasks.
Key findings
- MedSAM was trained on 1,570,263 image-mask pairs covering 10 imaging modalities and over 30 cancer types, and evaluated on 86 internal and 60 external segmentation tasks.
- On external validation, MedSAM achieved a median DSC of 87.8% (IQR 85.0–91.4%) for nasopharynx cancer segmentation, improving over SAM, specialist U-Net, and DeepLabV3+ by 52.3%, 15.5%, and 22.7%, respectively.
- MedSAM ranked first on most internal and external tasks, while SAM ranked last on almost all tasks; specialist models showed limited generalization on unseen targets.
- In a human annotation study, MedSAM-assisted annotation reduced time by 82.37% and 82.95% for two experts, and scaling training data from 10K to 100K to the full dataset improved performance.
Methods
- Curated public medical image segmentation datasets into 1,570,263 image-mask pairs across 10 modalities; images were normalized and resized to 1024 × 1024 × 3, with 3D CT/MR handled as 2D slices.
- MedSAM follows the SAM architecture with a ViT-Base image encoder, prompt encoder for bounding boxes, and mask decoder; it was initialized from SAM, fixed the prompt encoder, and trained the image encoder and mask decoder with combined dice and cross-entropy loss using AdamW, batch size 160, and 150 epochs on 20 A100 GPUs.
- Internal validation used an 80%/10%/10% training/tuning/validation split, while external validation datasets were held out; comparisons included SAM, modality-wise U-Net (nnU-Net), and DeepLabV3+ specialists, with DSC and normalized surface distance metrics and Wilcoxon signed-rank tests.
Limitations
- The paper states modality imbalance in the training set, with CT, MRI, and endoscopy dominating, potentially affecting less-represented modalities such as mammography.
- The paper states difficulty segmenting vessel-like branching structures because bounding box prompts can be ambiguous, e.g., arteries and veins sharing the same box in eye fundus images.
- Evident from the text: the human annotation study used only 10 adrenal tumor cases (733 slices) and two radiologists, so the reported time savings may not generalize broadly.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures





Verification
Claims the paper makes
7 Valid · 3 Not valid · 5 Not checked · 39 Own results and interpretations
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“Additionally, MedSAM provides a successful paradigm for adapting natural image foundation models to new domains, which can be further extended to biological image segmentation30, such as cell segmentation in light microscopy images31 and organelle segmentation in electron microscopy images32.”
§ Segment anything in medical images > Discussion
What the sources say
- Not checked — no full text of the sourcecites [30] Ma, J. & Wang, B. Towards foundation models of biological image segmentation. Nat. Method. 20, 953–955 (2023). source
- Could not verifycites [31] Ma, J. et al. The multi-modality cell segmentation challenge: towards universal solutions. Preprint at https://arxiv.org/abs/2308.05864 (2023). sourceExperimental — accuracy not yet measured
Our copy of Ma et al. (2024, Nature Methods) contains no mention of MedSAM, the Segment Anything Model, or natural-image foundation models anywhere (searched: 'MedSAM', 'Segment Anything Model SAM foundation model zero-shot medical image segmentation', 'adapting natural image foundation models to new domains'). The closest material is the Introduction and 'Existing SOTA cell segmentation algorithms' sections, which credit Cellpose 2.0 with 'the efficacy of adapting a pre-trained model to new images' via transfer learning, and the Discussion's 'successful proof of concept of generalist cell segmentation algorithms' — i.e., the paper does demonstrate generalist biological (microscopy) image segmentation, but never frames it as an extension of a MedSAM-style natural-image foundation-model paradigm. Nothing in our copy either supports or contradicts the attributed MedSAM claim. (LLM, conf=0.72)
- Not checked — no full text of the sourcecites [32] Xie, R., Pang, K., Bader, G.D., Wang, B. Maester: masked autoencoder guided segmentation at pixel resolution for accurate, self-supervised subcellular structure recognition. In IEEE Conference on Computer Vision and Pattern Recognition. 3292–3301 (IEEE, 2023). source
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“while existing deep learning-based interactive segmentation methods focus mainly on limited tasks and image modalities.”
§ Segment anything in medical images > Introduction
What the sources say
- Partly supportedcites [4] Wang, G. Deepigeos: a deep interactive geodesic framework for medical image segmentation. In IEEE Transactions on Pattern Analysis and Machine Intelligence 41, 1559–1572 (IEEE, 2018). sourceExperimental — accuracy not yet measured
The Related Works section describes deep-learning approaches for specific tasks, including semantic segmentation, fetal MRI segmentation, and interactive object selection. It does not establish that existing methods mainly focus on a limited range of tasks or image modalities; its stated criticism is that several methods do not allow interactive refinement at test time. (LLM, conf=0.74)
- Partly supportedcites [9] Wang, G. Interactive medical image segmentation using deep learning with image-specific fine tuning. In IEEE Transactions on Medical Imaging 37, 1562–1573 (IEEE, 2018). sourceExperimental — accuracy not yet measured
The Introduction says there are very few studies using CNNs for interactive segmentation and that current CNNs struggle to generalize to object classes absent from their training sets, supporting the limited-tasks point. It does not establish that existing interactive methods focus mainly on limited image modalities; its discussion of MRI applications concerns the proposed method. (LLM, conf=0.83)
- Partly supportedcites [10] Zhou, T. Volumetric memory network for interactive medical image segmentation. Med. Image Anal. 83, 102599 (2023). sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) says many deep interactive segmentation approaches focus only on 2D medical images, supporting a limitation in the kinds of images they handle. It does not say that existing methods focus mainly on a limited range of segmentation tasks or imaging modalities; its other stated limitation concerns the effort needed for user corrections. (LLM, conf=0.78)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The loss function is the unweighted sum between dice loss and cross-entropy loss, which has been proven to be robust in various segmentation tasks1.”
- Also stated in: § Loss function · plain
“We used the unweighted sum between cross-entropy loss and dice loss40 as the final loss function since it has been proven to be robust across different medical image segmentation tasks41.”
§ Segment anything in medical images > Methods > Training protocol and experimental setting
What the sources say
- Partly supportedcites [1] Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J. & Maier-Hein, K. H. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Method. 18, 203–211 (2021). sourceExperimental — accuracy not yet measured
Methods (p. 16) specifies a sum of cross-entropy and Dice loss, and B.1 Blueprint Parameters (p. 25) says the terms are averaged because combining them empirically improved training stability and segmentation accuracy. The paper reports strong nnU-Net results across diverse tasks (p. 6), but does not establish that this loss combination itself was proven robust across those tasks. (LLM, conf=0.85)
- Not checked — no full text of the sourcecites [41] Ma, J. Loss odyssey in medical image segmentation. Med. Image Anal. 71, 102035 (2021). source
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § Loss function · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In contrast, MedSAM maintained consistent performance across both internal and external validation sets.”
§ Segment anything in medical images > Methods > Training protocol and experimental setting
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“the performance adhered to the scaling rule, where increasing the number of training images significantly improved the performance in both internal and external validation sets.”
§ Segment anything in medical images > Results > The effect of training dataset size
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
25 more claims
Create a free account to see all 54 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 9
- It may state two separate claims as one. 18
- The quote only partly states this claim. 1
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“When compared with MedSAM53, a medical image segmentation foundation model, PanDerm showed slightly better accuracy (0.5% improvement, P = 0.025 and 0.112; Supplementary Table 22).”
From A Multimodal Vision Foundation Model for Clinical Dermatology (2025) · cites this paper as [53]
The cited source says what the claim attributes to it.
In the citing paper
“Adaptations of these models for the medical domain, such as MedSAM [15] and SAM-Med2D [4], have further demonstrated their potential to reduce annotation time.”
From Beyond Manual Annotation: A Human-AI Collaborative Framework for Medical Image Segmentation Using Only "Better or Worse" Expert Feedback (2025) · cites this paper as [15]
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 47,472 · cited 2×
“MedSAM, as the inaugural foundation model in medical image segmentation, holds great potential to accelerate the advancement of new diagnostic and therapeutic tools, and ultimately contribute to improved patient care [11].”
New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1)
Eisenhauer, E. A., Therasse, P., Bogaerts, J., Schwartz, L. H., Sargent, D., Ford, R., Dancey, J., Arbuck, S., Gwyther, S., Mooney, M., et al. (2009) · European Journal of Cancer · doi:10.1016/j.ejca.2008.10.026 · Cited by 29,534 · cited 1×
“The input image ( 1024 × 1024 × 3 ) was reshaped into a sequence of flattened 2D patches with the size 16 × 16 × 3 , yielding a feature size in image embedding of 64 × 64 after passing through the image encoder, which is 16 × downscaled.”
Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation
Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018) · arXiv · doi:10.48550/arxiv.1802.02611 · Cited by 20,063
Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation
Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-01234-2_49 · Cited by 20,054
Decoupled Weight Decay Regularization
Loshchilov, I., & Hutter, F. (2017) · arXiv · doi:10.48550/arxiv.1711.05101 · Cited by 12,522
Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., et al. (2023) · 2023 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.00371 · Cited by 11,953
V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation
Milletari, F., Navab, N., & Ahmadi, S.-A. (2016) · 2016 Fourth International Conference on 3D Vision (3DV) · doi:10.1109/3dv.2016.79 · Cited by 11,931
nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation
Isensee, F., Jaeger, P. F., Kohl, S. A. A., Petersen, J., & Maier-Hein, K. H. (2021) · Nature Methods · doi:10.1038/s41592-020-01008-z · Cited by 10,477 · cited 4×
“For each category, we trained a specialist model based on nnU-Net, which achieved SOTA performance on many segmentation tasks [3].”
BackgroundMethodInfluentialMasked Autoencoders Are Scalable Vision Learners
He, K., Chen, X., Xie, S., Li, Y., Dollar, P., & Girshick, R. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01553 · Cited by 8,489 · cited 1×
“The Wilcoxon signed-rank test results are marked on the DSC and NSD score tables (Supplementary Table 5-11).”
Research works printed but not linked: 1
- Vaswani, A. et al. Attention is all you need. In Advances in Neural Information Processing Systems, Vol. 30 (Curran Associates, Inc., 2017).
Cited by
Segment anything model for medical image analysis: An experimental study
Mazurowski, M. A., Dong, H., Gu, H., Yang, J., Konz, N., & Zhang, Y. (2023) · Medical Image Analysis · doi:10.1016/j.media.2023.102918 · Cited by 802
Segment anything model for medical images?
Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., et al. (2024) · Medical Image Analysis · doi:10.1016/j.media.2023.103061 · Cited by 536
Medical SAM adapter: Adapting segment anything model for medical image segmentation
Wu, J., Wang, Z., Hong, M., Ji, W., Fu, H., Xu, Y., Xu, M., & Jin, Y. (2025) · Medical Image Analysis · doi:10.1016/j.media.2025.103547 · Cited by 677
Chen, K., Liu, C., Chen, H., Zhang, H., Li, W., Zou, Z., & Shi, Z. (2024) · IEEE Transactions on Geoscience and Remote Sensing · doi:10.1109/tgrs.2024.3356074 · Cited by 383
Takahashi, S., Sakaguchi, Y., Kouno, N., Takasawa, K., Ishizu, K., Akagi, Y., Aoyama, R., Teraya, N., Bolatkan, A., Shinkai, N., et al. (2024) · Journal of Medical Systems · doi:10.1007/s10916-024-02105-8 · Cited by 343
“The second is the rapid development of foundation models that employ attention-based mechanisms, as exemplified by Meta’s SAM and MedSAM models [98, 99].”
Method
Metadata
- Authors
- Ma, J., He, Y., Li, F., Han, L., You, C., & Wang, B.
- Affiliations
- University Health Network (CA)
- Western University (CA)
- New York University (US)
- Yale University (US)
- Corresponding author
- Wang, Bo
- Venue
- Nature Communications
- Published
- 2024-01-22
- Volume
- 15
- Issue
- 1
- Pages
- 654-654
- ISSN
- 2041-1723 (ISSN-L)
- PMC
- PMC10803759
- OpenAlex
- W4391109864
- Semantic Scholar
- 9c3a827931fbd3cdb0620db3accd498548eb7bb2
- Cited by
- 2,913
- References
- 44
- Citations by year
- 2026: 1023 · 2025: 1233 · 2024: 612
- FWCI
- 622.56
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- Nature Communications (publishedVersion, cc-by)
- PubMed (publishedVersion)
- PubMed Central (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Canadian Network for Research and Innovation in Machining Technology, Natural Sciences and Engineering Research Council of Canada (RGPIN-2020-06189 and DGECR-2020-00294)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
- Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › COVID-19 diagnosis using AI
- Keywords
- medical image segmentation, MedSAM, foundation models, cross-modal generalization, external validation
- MeSH terms
- Diagnostic Imaging, Image Processing, Computer-Assisted
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash and gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Segment anything in medical images. https://app.cochat.ai/papers/10.1038/s41467-024-44824-z
- The paper
- Ma, J., He, Y., Li, F., Han, L., You, C., & Wang, B. (2024). Segment anything in medical images. Nature Communications. https://doi.org/10.1038/s41467-024-44824-z
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.