REVIEW 5 major objections 5 minor 2 references
SegmentAnyMuscle: A universal muscle segmentation model across different locations in MRI
T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A single deep-learning ensemble can segment muscle across all major MRI body locations and sequence types, reaching 88.45% Dice on common sequences and 86.21% on hard cases.
desk verdict A solid, reproducible engineering contribution—a broad MRI muscle segmentation model and dataset—whose absolute DSC numbers need an inter-rater baseline before being taken at face value. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the ensemble of two segmentation backbones: nnU-Net, a self-configuring U-Net pipeline that adapts preprocessing and architecture to the dataset, and a fine-tuned Segment Anything Model with adapter layers inserted at each transformer block, including a Mixture of Adapters variant that weights multiple adapters. The ensemble simply averages the two probability maps before thresholding. Equally central is the training-data strategy: 316 natively annotated MRIs across 11 body regions, augmented with 301 paired exams whose masks were transferred from aligned sequences after manual verification. This pairing is what gives the model its sequence diversity without proportional annotation cost.
What would settle it
Run the released weights on a multi-institutional MRI dataset with independently created ground truth using the same >50% muscle rule, and compare average Dice on T1, T2, and Dixon sequences per body region; if the common-sequence average falls well below 88.45% or varies strongly across institutions, the paper's universal applicability claim would not survive.
Extended reading notes
Core claim
On its own terms, the paper claims that a fully automated algorithm can perform binary muscle segmentation anywhere in the body on MRI. The authors constructed what they describe as one of the largest labeled MRI muscle datasets to date, with annotations covering chest, abdomen, thoracic and lumbar spine, shoulder, humerus, hip, thigh, knee, lower leg, and small structures of the hand, wrist, foot, and ankle. They extended the training set without extra annotation labor by transferring verified masks between aligned sequences of the same patient and view. The resulting ensemble of nnU-Net and fine-tuned SAM outperformed each component alone and beat standard baselines by 13.46 and 22.94 Dice points on the two test sets. Predicted skeletal muscle volume correlated with manual ground-truth volume at Pearson r >= 0.99, so the authors conclude the model is accurate enough for downstream quantitative muscle assessment.
Load-bearing premise
The entire accuracy story rests on the hand-built ground truth: any pixel that radiologists judged more than 50% likely to be muscle was labeled muscle, and transferred masks were trusted after visual alignment checks, so systematic annotation errors or inconsistent handling of fatty, atrophied, or tendon-rich muscle would directly change the reported Dice scores.
Editorial extensions
If this is right
- Automated muscle segmentation can be applied to routine MRI exams across 11 body regions, removing the 5–11 hours of manual annotation normally required per volume.
- Skeletal muscle volume can be estimated automatically with near-perfect correlation to manual measurements, supporting body-composition biomarkers such as skeletal muscle index.
- Performance degrades only mildly on rare sequences and abnormal cases, at 86.21% versus 88.45% Dice, so the model can still be used when atrophy, hardware artifacts, or noise are present.
- Because code and weights are public, other groups can reproduce the reported accuracy, retrain on local data, or apply the model to large population studies.
- The model can serve as an automated annotation aid for building multi-class muscle segmentations in future work.
Reading between the lines
- If the pairing strategy generalizes, it offers a blueprint for expanding segmentation datasets to new MRI sequences without collecting new ground truth, though it assumes patient-specific alignment is always verified by a human.
- The reported numbers are single-center; external data would likely reveal a performance gap, and the authors themselves call for multi-institutional validation before clinical use.
- The >50% muscle rule and inclusion of tendons in the muscle class mean the binary mask bakes in annotation subjectivity that will matter when the masks are used to quantify muscle quality, not just volume.
- The model's consistent behavior across locations and demographics suggests that a single universal muscle segmenter could be a reusable substrate for downstream tasks like sarcopenia screening and surgical planning, but those applications are not yet proven by this study.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents SegmentAnyMuscle, a deep-learning pipeline for binary muscle segmentation in MRI volumes across 11 body locations and 19 sequence types. A development set of 316 MRIs from 114 patients was built, including a mask-transfer pairing strategy to increase sequence diversity, and two held-out test sets were used: Test Set A (28 common-sequence MRIs) and Test Set B (18 challenging MRIs with less common sequences, atrophy, hardware, and noise). The authors compare nnU-Net, fine-tuned SAM, an ensemble of both, and several baselines, reporting an ensemble DSC of 88.45% on Test Set A and 86.21% on Test Set B, with statistically significant improvements over the individual models. They also show high correlation between predicted and ground-truth muscle volume. The code and trained weights are publicly released.
Significance. The paper's main contribution is a publicly available, multi-location MRI muscle segmentation model and a large annotated dataset, which are valuable resources for body-composition research. The evaluation on two independent, held-out test sets with deliberately challenging cases is a strength, as is the inclusion of multiple architectures and an ensemble baseline. The claim of 'universal' segmentation is, however, tempered by the single-institution data and the subjective annotation protocol; the reported accuracy numbers cannot be fully interpreted without a measure of human annotation variability. The public release of code and weights is a concrete step toward reproducibility.
major comments (5)
- [eMethod2.1, Results (Segmentation Accuracy)] The ground-truth annotation rule (a pixel is muscle if radiologists determine >50% probability of muscle) is subjective, and no inter-rater or intra-rater variability is reported. Since all DSC values (88.45% and 86.21%) are computed against these masks, the reported accuracy may be near or above the ceiling of human annotation agreement, or it may be inflated by systematic annotation bias. The authors should report inter-rater DSC (or another agreement metric) on a subset of the test volumes, ideally between the two senior radiologists, and compare model performance against that human benchmark. This is necessary to interpret the reported DSC as segmentation accuracy independent of the annotation process.
- [eMethod2.2, Results (Patient Characteristics), eTable1] There is an inconsistency in the reported number of paired/training volumes. The Results section states that the pairing strategy 'augmented the model development set with 301 additional exams,' while eMethod2.2 states that 145 MRI exams were added and that the training set increased from 145 to 290 volumes; eTable1 confirms 290 training volumes and 26 validation volumes (total 316). The number 301 is therefore inconsistent with the supplementary description and the tabulated data. Correct this discrepancy, as it affects the reported dataset composition and reproducibility.
- [Results (Segmentation Accuracy), Figures 4 and 5] DSC and HSS are reported only as means over 28 and 18 test volumes, without confidence intervals, standard deviations, or per-volume distributions. With these small sample sizes, the claimed statistically significant improvement of the ensemble over its components may be driven by a few cases. Provide per-subject results (e.g., boxplots or scatter plots) and confidence intervals for the mean DSC, and clarify whether the one-sided Wilcoxon signed-rank test was pre-specified and what the effect size is.
- [Methods (Statistical Analysis), Methods (Estimation of Skeletal Muscle Index), Results (Skeletal Muscle Mass…] The Methods state that 'the association of predicted SMM with the ground truth SMM was evaluated using univariable logistic regression analyses,' but the Results report Pearson correlation coefficients (r >= 0.99, P < 0.001), which is a different analysis. Additionally, the text defines SMI as SMV divided by height, but the reported analysis appears to use SMV volumes across multiple locations without explicitly computing SMI. Clarify which quantity was actually computed (SMV or SMI), which statistical test was used, and provide the regression details if logistic regression was indeed performed.
- [Methods (Segmentation Model Architectures), Results] The 'Mixture of Adapters' method is described as a proposed approach, but the Results section only reports the ensemble of nnU-Net and fine-tuned SAM; it is unclear whether Mixture of Adapters was used as the fine-tuned SAM variant, tested separately, or omitted. Specify which model configurations were evaluated and include Mixture of Adapters results if it was run, or state explicitly that it was not used in the final evaluation.
minor comments (5)
- [Abstract, Discussion] The abstract and discussion describe the model as 'universal' and applicable across 'diverse settings,' but the evaluation is exclusively from a single institution (Duke University Health System) and excludes head/face regions. Consider softening the wording to 'multi-location' and explicitly state the single-institution limitation in the abstract.
- [eMethod2.1] The annotation definition includes tendons as muscle. This is a nonstandard anatomical convention and may affect the comparability of the reported DSC with other muscle segmentation studies. State this choice in the main text.
- [Results (Segmentation Accuracy)] The text says 'Both models have outperformed the remaining competing methods by a large margin' but does not provide statistical comparisons against UNet, AttentionUNet, or SwinUNet. Report pairwise significance tests or confidence intervals for these comparisons, or explicitly state that no formal comparison was made.
- [Introduction, Bibliography] Reference 26 is a web link; consider citing a peer-reviewed article for the skeletal muscle mass index definition.
- [eMethod1] The supplementary method describes sequence categorization via keyword detection; it would be helpful to mention how ambiguous or unlisted sequence descriptions (e.g., 'unknown') were handled, since 15 training volumes are labeled 'unknown' in eTable1.
Circularity Check
No significant circularity: the model's accuracy is evaluated on independent test sets against externally defined ground truth, and no predicted quantity reduces by construction to a fitted input.
full rationale
The paper's central claim is an empirical segmentation performance result: an ensemble of nnU-Net and fine-tuned SAM achieves 88.45% DSC on Test Set A and 86.21% DSC on Test Set B. These test sets are explicitly described as independent from the model development set (Figure 1, eTable1), and the Dice metric is computed against manually created ground-truth masks, not against the model's own outputs or any parameter fitted to the test data. The skeletal muscle index correlation (Pearson r >= 0.99) is a validation of predicted volume against ground-truth volume, not a derivation of the prediction from the ground truth; high correlation is a downstream consequence of high segmentation accuracy, not a definitional identity. The self-citations in the paper are not load-bearing: reference 12 (SegmentAnyBone) is cited only as related prior work in the introduction, and reference 21 is cited alongside an external reference (20) for the SAM adapter architecture. Neither citation is used to justify the central performance claim or to forbid alternative approaches. The acknowledged limitation of single-institution data and the absence of inter-rater variability measurements are correctness and generalizability concerns, not circularity: they do not make the reported DSC numbers equivalent to the model's training inputs by construction. Therefore, no circular step is present, and the appropriate score is 0.
Assumptions & free parameters
free parameters (1)
- Muscle pixel probability threshold =
50%
assumptions (3)
- domain assumption Manual annotations are an accurate ground truth for muscle tissue.
- domain assumption Annotation masks can be transferred across aligned sequences without modification.
- domain assumption The single-center Duke dataset is representative enough to support the 'universal' claim.
Cite this review
Pith. "Pith review of SegmentAnyMuscle: A universal muscle segmentation model across different locations in MRI." pith.science (2026). https://pith.science/paper/AKABTU4M
@misc{pith2026250622467,
author = {Pith},
title = {Pith review of: SegmentAnyMuscle: A universal muscle segmentation model across different locations in MRI},
year = {2026},
howpublished = {\url{https://pith.science/paper/AKABTU4M}},
note = {Machine review of arXiv:2506.22467}
}
read the original abstract
The quantity and quality of muscles are increasingly recognized as important predictors of health outcomes. While MRI offers a valuable modality for such assessments, obtaining precise quantitative measurements of musculature remains challenging. This study aimed to develop a publicly available model for muscle segmentation in MRIs and demonstrate its applicability across various anatomical locations and imaging sequences. A total of 362 MRIs from 160 patients at a single tertiary center (Duke University Health System, 2016-2020) were included, with 316 MRIs from 114 patients used for model development. The model was tested on two separate sets: one with 28 MRIs representing common sequence types, achieving an average Dice Similarity Coefficient (DSC) of 88.45%, and another with 18 MRIs featuring less frequent sequences and abnormalities such as muscular atrophy, hardware, and significant noise, achieving 86.21% DSC. These results demonstrate the feasibility of a fully automated deep learning algorithm for segmenting muscles on MRI across diverse settings. The public release of this model enables consistent, reproducible research into the relationship between musculature and health.
Reference graph
Works this paper leans on
-
[4]
Pickhardt PJ, Graffy PM, Zea R, et al. Automated CT biomarkers for opportunistic prediction of future cardiovascular events and mortality in an asymptomatic screening population: a retrospective cohort study. Lancet Digit Health. 2020;2(4):e192-e200. doi:10.1016/S2589-7500(20)30025-X 5. Erlandson MC, Lorbergs AL, Mathur S, Cheung AM. Muscle analysis using...
-
[21]
MRI hip left with and without contrast
Gu H, Dong H, Yang J, Mazurowski MA. How to build the best medical image segmentation algorithm using foundation models: a comprehensive empirical study with Segment Anything Model. Published online May 13, 2024. doi:10.48550/arXiv.2404.09957 22. Wang Y, Agarwal S, Mukherjee S, et al. AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning. Pu...
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.