CoChat
Sign Up
Cited by 309Medical Image Segmentation TechniquesMedical Imaging and AnalysisBrain Tumor Detection and Classification

MUSE: MUlti-atlas region Segmentation utilizing Ensembles of registration algorithms and parameters, and locally optimal atlas selection

Open access
Doshi, Jimit, Erus, Guray, Ou, Yangming, Resnick, Susan M., Gur, Ruben C., Gur, Raquel E., Satterthwaite, Theodore D., Furth, Susan +1 more·2016·NeuroImage
DOI 10.1016/j.neuroimage.2015.11.073PMC PMC4806537
Open this paper in CoChat — free account

Abstract

Atlas-based automated anatomical labeling is a fundamental tool in medical image segmentation, as it defines regions of interest for subsequent analysis of structural and functional image data. The extensive investigation of multi-atlas warping and fusion techniques over the past 5 or more years has clearly demonstrated the advantages of consensus-based segmentation. However, the common approach is to use multiple atlases with a single registration method and parameter set, which is not necessarily optimal for every individual scan, anatomical region, and problem/data-type. Different registration criteria and parameter sets yield different solutions, each providing complementary information. Herein, we present a consensus labeling framework that generates a broad ensemble of labeled atlases in target image space via the use of several warping algorithms, regularization parameters, and atlases. The label fusion integrates two complementary sources of information: a local similarity ranking to select locally optimal atlases and a boundary modulation term to refine the segmentation consistently with the target image's intensity profile. The ensemble approach consistently outperforms segmentations using individual warping methods alone, achieving high accuracy on several benchmark datasets. The MUSE methodology has been used for processing thousands of scans from various datasets, producing robust and consistent results. MUSE is publicly available both as a downloadable software package, and as an application that can be run on the CBICA Image Processing Portal (https://ipp.cbica.upenn.edu), a web based platform for remote processing of medical images.

AI snapshot

Written from the paper’s full text.

TL;DR

MUSE is a multi-atlas segmentation framework that builds a broad ensemble of warped atlas labels using multiple atlases, registration algorithms, and regularization parameters, then fuses them with local similarity ranking and boundary modulation. It consistently outperformed single registration method/parameter ensembles and achieved the highest average Dice score in the MICCAI 2013 mid-brain segmentation challenge.

Key findings

  • The complete ensemble of warps consistently outperformed ensembles from a single registration algorithm on each dataset, with significant differences (p < 0.01, paired t-test); adding local similarity ranking significantly increased performance over simple majority voting (p < 0.01 for all datasets).
  • Full MUSE with boundary modulation improved segmentation accuracy for all datasets except NIREP; differences were significant except for OASIS (p = 0.06185). In the MICCAI 2013 challenge, MUSE obtained the highest average Dice score (d = 0.8686) in the mid-brain segmentation category.
  • Segmentation accuracy consistently increased with more atlases until reaching a stable value around 7 atlases. On a pooled multi-site dataset of clinically normal subjects (n = 1029, age range 8 to 94), MUSE obtained the highest brain-age prediction accuracy among MUSE, STAPLE, and Joint Label Fusion in terms of Pearson’s r, concordance correlation coefficient, and mean absolute error.
  • MUSE has been used for processing thousands of scans from various datasets, producing robust and consistent results, and is publicly available as a downloadable package and via the CBICA Image Processing Portal.

Methods

  • MUSE constructs an ensemble of k warped atlas label maps by varying the atlas, deformable registration algorithm, and smoothness/regularization parameter. The paper used DRAMMS with g = {0.1, 0.2} and ANTS with s = {0.25, 0.5}; label fusion integrates a local similarity rank score based on a Gabor-filter attribute descriptor and boundary modulation terms b1 and b2.
  • Internal validation used four public benchmark datasets with expert reference labels: BrainWeb, IBSR, NIREP, and OASIS. Leave-one-out cross-validation was applied independently per dataset, and Dice similarity coefficient was used for quantitative evaluation. The MICCAI 2013 challenge evaluation used the OASIS dataset for mid-brain structures.
  • Multi-site experiments pooled clinically normal subjects from BLSA, ADNI-1, PNC, BBL-NC, and NiCK (n = 1029). For each subject, ROI segmentation was performed with MUSE, STAPLE, and Joint Label Fusion using 11 atlases selected from 35 OASIS atlases; ROI volumes were then used in a supervised learning framework to predict brain age with leave-one-out cross-validation, reporting Pearson’s r, concordance correlation coefficient, and mean absolute error.

Limitations

  • The paper states that one limitation of MUSE is its high computational requirement, a general problem for multi-atlas segmentation methods; this is mitigated by parallelizable registrations, multi-scale registration implementations, and user control over the number of registrations.
  • The paper states that for NIREP, the boundary modulation term did not improve quantitative results, which the authors attribute to low accuracy of the reference ROI masks, particularly under-segmented deep brain structures.
  • Evident from the text: direct quantitative evaluation of multi-site segmentation accuracy was not possible because common reference labels for multiple datasets were unavailable, so multi-site robustness was assessed indirectly through brain-age prediction.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.neuroimage.2015.11.073

Cites

55 of 55 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

2 of 9 cited sources checked · 1 supported · 0 flagged · accuracy not yet measured

  • Findingown data

    The additional boundary modulation term results in higher segmentation accuracy for all datasets, except NIREP.

    “The additional boundary modulation term results in higher segmentation accuracy for all datasets, except NIREP.”
    Quote in the paperSubstantial claim
  • Findingown data

    The addition of the similarity based weighting term significantly increases the performance compared to simple majority voting (p < 0.01 for all datasets).

    “The addition of the similarity based weighting term significantly increases the performance compared to simple majority voting (p < 0.01 for all datasets).”
    Quote in the paperSubstantial claim
  • Findingnone given

    MUSE achieved consistently high Dice scores for the segmentation of important deep brain structures, such as hippocampus, thalamus and caudate.

    “MUSE achieved consistently high Dice scores for the segmentation of important deep brain structures, such as hippocampus, thalamus and caudate”
    Quote in the paperSubstantial claim
  • Findingown data

    Differences between using the complete ensemble of warps versus any set of warps from a single registration algorithm are significant as well.

    “Differences between using the complete ensemble of warps versus any set of warps from a single registration algorithm are significant as well.”
    Quote in the paperSubstantial claim
  • Findingown data

    The segmentation accuracy consistently increases with more atlases until it reaches a stable value around 7 atlases.

    “We observe that the segmentation accuracy consistently increases with more atlases until it reaches a stable value around 7 atlases (Figure 5).”
    Quote in the paperSubstantial claim
  • May combine two claims. 18
  • Classified as probably not a substantial claim. 13

  • May combine two claims.

    Premisecitation

    More aggressive deformations may result in higher atlas-to-target regional overlaps, which is not ideal for registration but preferable for atlas-based segmentation.

    “In comparative survey papers (Ou et al., 2014; Klein et al., 2009) and in many other studies, it was reported that more aggressive deformations may result in higher atlas-to-target regional overlaps, which is not ideal for registration but preferable for atlas-based segmentation.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Ou et al., 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Klein et al., 2009 source

      The Discussion (page unavailable) reports a modest correlation between a deformation’s degrees of freedom and registration accuracy, consistent with more flexible methods achieving better measured results. The Caveats section (page unavailable) warns that overlap measures can miss misregistration and folding. The paper does not establish that aggressive deformations are preferable for atlas-based segmentation but not ideal for registration. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Deformable image registration is a core component of all multi-atlas segmentation (MAS) methods, and the quality of individual registrations has a very high impact on the accuracy of the final segmentation.

    “deformable image registration (Sotiras et al., 2013) is a core component of all MAS methods, and the quality of individual registrations has a very high impact on the accuracy of the final segmentation.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Sotiras et al., 2013 source

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Imaging features derived using MUSE segmentation could accurately predict brain age.

    “Imaging features derived using MUSE segmentation could accurately predict the brain age”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In internal validation experiments, the similarity ranking and boundary modulation terms contributed to segmentation performance.

    “In our internal validation experiments we demonstrated the contribution of the similarity ranking and the boundary modulation terms to segmentation performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For each dataset, the complete ensemble of warps consistently outperforms ensembles of warps from a single registration algorithm, with significant differences (p < 0.01 with a paired t-test).

    “We observe that, for each dataset, the complete ensemble of warps consistently outperforms ensembles of warps from a single registration algorithm, with significant differences (p < 0.01 with a paired t-test).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    The scores of the three methods with the highest ranks in the challenge were similar and comparable.

    “However, the scores of the three methods with the highest ranks were similar and comparable.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    MUSE obtained the highest average Dice score (d = 0.8686) in the mid-brain segmentation category in the MICCAI 2013 Challenge Workshop on Segmentation, and maintained first rank as of 11/02/2015.

    “MUSE obtained the highest average Dice score (d = 0.8686) in the mid-brain segmentation category, and it maintains the first rank as of 11/02/20153.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    MUSE has been used to process thousands of scans from various datasets and produces robust and consistent results.

    “The MUSE methodology has been used for processing thousands of scans from various datasets, producing robust and consistent results.”
    • Also stated in: § 1. Introduction · boosted
      “we applied MUSE on thousands of images from several large-scale neuroimaging studies, and showed the robustness of our method and consistency of segmentations for datasets with significant differences in scanner characteristics and sample demographics, by accurately estimating brain age from segmented ROIs.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    MUSE performed with high accuracy on the testing dataset and obtained the highest average Dice score in the MICCAI 2013 segmentation challenge.

    “Consistently with our cross-validated segmentation accuracy on the training dataset, MUSE performed with high accuracy on the testing dataset and obtained the highest average Dice score.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The MUSE ensemble approach consistently outperforms segmentations using individual warping methods alone, achieving high accuracy on several benchmark datasets.

    “The ensemble approach consistently outperforms segmentations using individual warping methods alone, achieving high accuracy on several benchmark datasets.”
    • Also stated in: § 1. Introduction · boosted
      “we confirmed that the ensemble approach consistently outperforms segmentations obtained using individual warping methods/parameters alone.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    We trained an ensemble of regressors on the ROI volumes using the complete dataset (n = 1029) with leave-one-out cross-validation, and predicted the age of each subject using the trained models.

    “We trained an ensemble of regressors on the ROI volumes using the complete dataset (n = 1029) with leave-one-out cross-validation, and predicted the age of each subject using the trained models.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Anatomies closer to an atlas are well represented by a diffeomorphism.

    “Anatomies closer to an atlas are well represented by a diffeomorphism.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The ensemble approach efficiently solves the problem of selecting optimal algorithm/parameter values by covering a range of possible values without having to pick one, and provides a robust segmentation tool.

    “As stated in Rohlfing and Maurer (2005), the ensemble approach efficiently solves this problem by covering a range of possible values without having to pick one, and in this way provides a robust segmentation tool.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Rohlfing and Maurer (2005) source

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    The general practice in current MAS methods is to select a single warping algorithm and to use it with a single set of registration parameters.

    “the general practice is to select a single warping algorithm and to use it with a single set of registration parameters.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    A single fold cross-validation was applied for each dataset to segment ROIs in each image, preferred in order not to over-tweak parameters for a specific reference atlas dataset and to keep the method more generalizable to unseen datasets.

    “A single fold cross-validation was applied for each dataset to segment ROIs in each image. This was preferred in order not to over-tweak the parameters for a specific set of reference atlas dataset and thus to keep our method more generalizable to unseen datasets.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    All experiments were performed using leave-one-out cross validation, applying it independently for each dataset.

    “All experiments were performed using leave-one-out cross validation, applying it independently for each dataset.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    In Rohlfing and Maurer (2005), classifier combinations consistently improved classification accuracy, and improvements in accuracy were possible with various parameterizations of the non-rigid registration technique, even using a single atlas.

    “the authors observed that in all cases classifier combinations consistently improved classification accuracy, and that improvements in accuracy were possible with various parameterizations of the non-rigid registration technique, even using a single atlas.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Rohlfing and Maurer (2005) source

  • May combine two claims.

    Interpretationreasoning

    Each method and parameter set can be relatively less or more accurate in certain areas of the brain, thereby rendering an ensemble-based segmentation advantageous.

    “Each method and parameter set can be relatively less or more accurate in certain areas of the brain, thereby rendering an ensemble-based segmentation advantageous.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Each method and parameter set can be relatively less or more accurate in certain areas of the brain.

  • Classified as probably not a substantial claim.

    Premisecitation

    In Bai et al. (2012), image registration plays a more crucial role in segmentation compared to the complexity of the segmentation model.

    “the authors investigated the roles of image registration and segmentation model complexity for mouse brain segmentation using 4 different registration algorithms, and concluded that image registration plays a more crucial role in segmentation compared to the complexity of the segmentation model.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Bai et al. (2012) source

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Anatomical correspondence may not be uniquely determined from intensity-based image attributes, which drive deformable registration algorithms.

    “anatomical correspondence may not be uniquely determined from intensity-based image attributes, which drive deformable registration algorithms.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    The intensity term makes the final segmentation consistent with the intensity profile of the target image.

    “The main purpose of the intensity term is to make the final segmentation consistent with the intensity profile of the target image.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: ROIs belong to a single tissue type with a smooth intensity variation.
    • The authors state this as an assumption: The intensity distribution of each ROI can be modeled as a normal distribution.

  • May combine two claims.

    Interpretationown data

    The ensemble of a multitude of warps, particularly using appropriate techniques for fusing them together, has significantly improved the segmentation accuracy, and provided robust segmentations.

    “We demonstrated that the ensemble of a multitude of warps, particularly using appropriate techniques for fusing them together, has significantly improved the segmentation accuracy, and provided robust segmentations.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Different registration methods provide complementary information about the anatomy.
    • Rests on an assumption: Each method and parameter set can be relatively less or more accurate in certain areas of the brain.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    A learning based approach is expected to improve the segmentation accuracy within a single dataset, but may also overfit to a specific set of reference labels and thus may have lower generalizability to new datasets, compared to a pure label fusion based approach without learning.

    “A learning based approach is expected to improve the segmentation accuracy within a single dataset, however it may also overfit to a specific set of reference labels and thus may have lower generalizability to new datasets, compared to a pure label fusion based approach without learning.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The use of multiple registration methods and atlases would linearly increase the computation time, but this is not a hard constraint for the practical application of MUSE because registrations are efficiently parallelizable and multi-scale implementation can generate warps of various smoothness levels for no additional cost.

    “While the use of multiple registration methods and atlases would linearly increase the computation time, this is not a hard constraint for the practical application of MUSE, because the registrations are very efficiently parallelizable, and notably, multi-scale implementation of registration algorithms can effectively generate warps of various smoothness levels for no additional cost.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Our results showed that this approach outperforms methods that are based on single parameter sets and registration algorithms, and can therefore provide a foundation for robust segmentation.

    “Our results showed that this approach outperforms methods that are based on single parameter sets and registration algorithms, and can therefore provide a foundation for robust segmentation.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Different registration methods provide complementary information about the anatomy.
    • Rests on an assumption: Each method and parameter set can be relatively less or more accurate in certain areas of the brain.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    One of the limitations of MUSE is the high computational requirement, which is a general problem for MAS methods.

    “One of the limitations of MUSE is the high computational requirement, which is a general problem for MAS methods.”
    Quote in the paperNot a substantial claim
26 of 26

  • The authors state this as an assumption.

    ROIs belong to a single tissue type with a smooth intensity variation.

    “with the assumption that the ROIs belong to a single tissue type with a smooth intensity variation.”

  • The authors state this as an assumption.

    The intensity distribution of each ROI can be modeled as a normal distribution.

    “The intensity distribution of each ROI is modeled as a normal distribution”

  • The paper relies on this without support.

    Different registration methods provide complementary information about the anatomy.

    “Different registration methods, which generally use notably different image features, energy formulations and optimization algorithms, provide complementary information about the anatomy.”

  • The paper relies on this without support.

    Each method and parameter set can be relatively less or more accurate in certain areas of the brain.

    “Each method and parameter set can be relatively less or more accurate in certain areas of the brain”
4 of 4

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?

  • Data from prior OASIS projects have been extensively used in the development of neuroimaging tools, models of healthy aging and AD timecourse, hypothesis-driven questions, and educational/training purposes.

    “Data from these projects have been extensively used in the development of neuroimaging tools, such as brain atlases 3,4 , tissue segmentation algorithms 5 , and de-identification tools 6 , used to create models of healthy aging 7,8 and AD timecourse 7-11 , used to answer hypothesis driven questions 12,13 , and used for educational and training purposes.”

    From OASIS-3: Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease (2019) · cites this paper as [5]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Cross-validation is described as both single-fold and leave-one-out

    Minor

    The methods say a single fold was used for each dataset, but the results say all experiments used leave-one-out cross-validation. Leave-one-out requires a separate fold for each image, rather than a single fold.

1 of 1

Cites

Flagged first, then by citations

1 of the 56 references couldn’t be matched to a known paper.

Cited by

Sorted by citations

Related

Authors
Doshi, J., Erus, G., Ou, Y., Resnick, S. M., Gur, R. C., Gur, R. E., Satterthwaite, T. D., Furth, S., & Davatzikos, C.
Affiliations
  • University of Pennsylvania (US)
  • Harvard University (US)
  • National Institute on Aging (US)
  • Children's Hospital of Philadelphia (US)
ORCID
Venue
NeuroImage
Publisher
Elsevier BV
Published
2016-02
Volume
127
Pages
186-195
ISSN
1053-8119 (ISSN-L)
DOI
10.1016/j.neuroimage.2015.11.073
PMC
PMC4806537
OpenAlex
W2196038229
Semantic Scholar
2cee0b733a1a63fe50ade982d837f892a00ea1c0
Cited by
309
References
56
Citations by year
2026: 26 · 2025: 45 · 2024: 35 · 2023: 41 · 2022: 39 · 2021: 34 · 2020: 27 · 2019: 25 · 2018: 15 · 2017: 12 · 2016: 10
FWCI
10.62
Type
journal article
Language
en
Open access
Licence
www.elsevier.com/tdm/userlicense/1.0/ (tdm)
Locations
  • NeuroImage (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion)
Funders
  • Intramural Research Program, National Institute on Aging, NIH
  • National Institutes of Health (R01-AG014971)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
  • Physical Sciences › Engineering › Biomedical Engineering › Medical Imaging and Analysis
  • Life Sciences › Neuroscience › Neurology › Brain Tumor Detection and Classification
Keywords
multi-atlas segmentation, anatomical labeling, atlas selection, label fusion, brain segmentation, medical image segmentation, web-based image processing
MeSH terms
Algorithms, Anatomy, Artistic, Atlases as Topic, Brain, Brain Mapping, Humans, Image Processing, Computer-Assisted, Pattern Recognition, Automated

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-18.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-18.

Cited sources are checked by 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