pith. sign in

arxiv: 2510.26635 · v3 · pith:XQCJ7EWPnew · submitted 2025-10-30 · 📡 eess.IV · cs.CV

SAMRI: Segment Any MRI

Zhao Wang , Wei Dai , Thuy Thanh Dao , Steffen Bollmann , Hongfu Sun , Craig Engstrom , Shekhar S. Chandra This is my paper

Pith reviewed 2026-05-21 20:39 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords MRI segmentationSegment Anything Modelfoundation model adaptationzero-shot generalizationprompt-based segmentationwhole-body MRIdecoder fine-tuningmedical image analysis
0
0 comments X

The pith

Fine-tuning only the mask decoder of SAM on over a million MRI slices yields more accurate whole-body segmentation than prior adaptations, especially for small structures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to demonstrate that an MRI-specific version of the Segment Anything Model can be built efficiently by updating just the mask decoder while leaving the image encoder unchanged from its natural-image pretraining. This matters to a sympathetic reader because full retraining of such models is expensive in time and compute, yet the targeted change still delivers higher accuracy on many anatomical targets across different MRI contrasts and shows good results on completely new datasets. The approach uses a massive collection of training examples covering dozens of structures and relies on simple box or point prompts to guide the output. If the central result holds, it points toward a practical way to turn general foundation models into reliable tools for medical image analysis without needing enormous resources for each new modality.

Core claim

SAMRI is produced by fine-tuning solely the mask decoder of the original SAM ViT-B/16 model on 1.1 million 2D MRI slice-mask pairs drawn from 30 datasets that together cover 47 targets, multiple contrasts including T1, T2, FLAIR and DWI, and whole-body anatomy. With focal-Dice loss and prompts consisting of bounding boxes plus optional points, the resulting model records a mean Dice similarity coefficient of 0.87 across the targets, outperforming the earlier MedSAM adaptation by 17.6 percent with the largest relative gains on small and medium-sized structures, while also performing well on six separate zero-shot test sets and requiring only modest memory at inference time.

What carries the argument

Decoder-only fine-tuning of the Segment Anything Model on a large MRI-specific corpus of slice-mask pairs, guided by box and point prompts.

If this is right

  • Rapid prompt-driven annotation becomes feasible for whole-body MRI on standard hardware using roughly 4.5 GB of memory.
  • Zero-shot performance on unseen MRI datasets reaches a mean Dice of 0.85 without additional training.
  • Training completes with 94 percent less time and 96 percent fewer trainable parameters than updating the entire model.
  • Gains are largest for structures whose masks occupy less than 3.5 percent of the image area.
  • The same decoder update strategy supports interactive use across any new MRI annotation task.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Pretrained natural-image features appear sufficient for MRI when the decoder is allowed to learn domain-specific mapping from prompts to masks.
  • Comparable decoder-only adaptation could be tested on CT or PET volumes to check whether the efficiency pattern generalizes across modalities.
  • Combining the model with simple 3D post-processing might extend accurate segmentation from individual slices to full volumetric MRI studies.
  • Clinical annotation pipelines could shift toward interactive prompting rather than exhaustive manual outlining for low-prevalence or small lesions.

Load-bearing premise

The image encoder pretrained on ordinary photographs continues to supply useful features for MRI despite differences in tissue contrast, intensity patterns, and artifacts.

What would settle it

Evaluating SAMRI on a fresh set of MRI scans containing many small targets and obtaining Dice scores that fall below those of a fully fine-tuned model or show no advantage over MedSAM would undermine the central claim.

Figures

Figures reproduced from arXiv: 2510.26635 by Craig Engstrom, Hongfu Sun, Shekhar S. Chandra, Steffen Bollmann, Thuy Thanh Dao, Wei Dai, Zhao Wang.

Figure 1
Figure 1. Figure 1: SAM family and the SAMRI training pipeline [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance vs. object size. (a) Median Dice (±IQR) by object-size bin; the horizontal dashed line marks the 0.80 as a heuristic consistent. The vertical dashed line separates large objects (>3.5% of image area) from small/medium (<3.5%). SAMRI attains the highest Dice across sizes, with the largest gains in the small–medium regime. (b) Histogram of training samples by object size (same x-axis), showing a … view at source ↗
read the original abstract

Summary: SAMRI is an MRI-specialized adaptation of the Segment Anything Model achieving superior whole-body MRI segmentation, particularly for small and clinically critical structures, through box and point prompts for rapid annotation. Purpose: Existing SAM adaptations treat MRI as a generic modality, overlooking variable tissue contrast, intensity inhomogeneity, and clinically important small structures. We propose an MRI-specialized foundation model with strong whole-body segmentation and zero-shot generalization for direct use on any MRI annotation task. Methods: SAMRI fine-tunes only the mask decoder of SAM (ViT-B/16), keeping encoders frozen to preserve pretrained representations and eliminate redundant passes-reducing training time by 94%, trainable parameters by 96%, and FLOPs by ~99% versus full-model retraining. Training used 1.1 million 2D slice-mask pairs from 30 datasets spanning 47 targets, T1/T2/FLAIR/DWI contrasts, and whole-body anatomy, with focal-Dice loss and bounding-box (with optional point) prompts. Sizes were stratified by mask area (small: <0.5%; medium: 0.5-3.5%; large: >3.5%), and significance assessed by the Wilcoxon signed-rank test. Results: SAMRI with box+point prompts achieved mean DSC 0.87 +/- 0.11 across 47 targets, outperforming MedSAM (0.74 +/- 0.24) by 17.6% (p < 0.05), with largest gains for small (+42.4%) and medium (+26.9%) structures. On six zero-shot datasets, SAMRI achieved mean DSC 0.85, outperforming baselines. Inference requires only ~4.5 GB VRAM through an interactive interface on standard hardware. Conclusion: Decoder-only fine-tuning on a large, MRI-specific corpus delivers superior whole-body segmentation with strong zero-shot generalization, particularly for small and clinically salient structures. Public code, pretrained models, and an interactive interface make SAMRI deployable for MRI segmentation research and clinical workflows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces SAMRI, an adaptation of the Segment Anything Model (SAM) for MRI by fine-tuning only the mask decoder (ViT-B/16 encoder kept frozen) on 1.1 million 2D slice-mask pairs from 30 datasets spanning 47 targets across T1/T2/FLAIR/DWI contrasts. It reports mean DSC of 0.87 ± 0.11 with box+point prompts, outperforming MedSAM (0.74 ± 0.24) by 17.6% (p < 0.05), with largest gains on small (+42.4%) and medium (+26.9%) structures, plus mean DSC 0.85 on six zero-shot datasets. The approach emphasizes efficiency (96% fewer trainable parameters, 94% less training time) and provides public code, models, and an interactive interface.

Significance. If the reported performance holds, SAMRI demonstrates that decoder-only fine-tuning on a large MRI-specific corpus can deliver strong whole-body segmentation with particular advantages for small and clinically relevant structures, while achieving substantial computational savings. The public release of code, pretrained models, and an interactive interface supporting ~4.5 GB VRAM inference is a clear strength that facilitates reproducibility and adoption in research and clinical workflows.

major comments (2)
  1. [Methods] Methods: The decision to freeze the pretrained natural-image ViT-B/16 encoder is justified by preservation of representations and efficiency gains, but no ablation is reported that compares this to a fine-tuned encoder on the same 1.1 M slices. This is load-bearing for the central claim of superior small-structure performance (DSC gains of +42.4%), as MRI domain shifts (contrast, bias fields, artifacts) could degrade encoder features without adaptation.
  2. [Abstract and Methods] Abstract and Methods: Exact train/validation/test splits across the 30 datasets, patient-level partitioning, and explicit checks for data leakage are not described. This detail is required to substantiate the zero-shot generalization results (mean DSC 0.85 on six held-out datasets) and the cross-dataset claims.
minor comments (2)
  1. [Results] The size stratification thresholds (<0.5%, 0.5-3.5%, >3.5%) are useful but would benefit from explicit justification or sensitivity analysis in the results section.
  2. [Results] Table or figure captions should clarify whether the reported standard deviations are across targets or across datasets.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive overall assessment of our work. We address each major comment in detail below, providing clarifications and indicating revisions that will be incorporated into the next version of the manuscript.

read point-by-point responses

  1. Referee: [Methods] Methods: The decision to freeze the pretrained natural-image ViT-B/16 encoder is justified by preservation of representations and efficiency gains, but no ablation is reported that compares this to a fine-tuned encoder on the same 1.1 M slices. This is load-bearing for the central claim of superior small-structure performance (DSC gains of +42.4%), as MRI domain shifts (contrast, bias fields, artifacts) could degrade encoder features without adaptation.

    Authors: We appreciate the referee pointing out the value of an ablation comparing frozen versus fine-tuned encoder configurations. Our choice to keep the ViT-B/16 encoder frozen was driven by two considerations: (1) preserving the rich, general visual features learned from large-scale natural-image pretraining, which prior work has shown transfer effectively to medical domains for low- and mid-level representations, and (2) realizing the reported efficiency gains (96% fewer trainable parameters, 94% less training time). Performing a full encoder fine-tuning ablation on the entire 1.1 M slice corpus would have required substantially greater compute resources and time, which was outside the scope of demonstrating an efficient decoder-only adaptation. In the revised manuscript we have expanded the Methods section with additional justification and supporting citations, and we have added a brief limitations paragraph in the Discussion noting that a controlled encoder-ablation study remains valuable future work. We believe these changes adequately address the concern while preserving the paper’s focus on practical, resource-efficient adaptation. revision: yes

  2. Referee: [Abstract and Methods] Abstract and Methods: Exact train/validation/test splits across the 30 datasets, patient-level partitioning, and explicit checks for data leakage are not described. This detail is required to substantiate the zero-shot generalization results (mean DSC 0.85 on six held-out datasets) and the cross-dataset claims.

    Authors: We agree that precise documentation of data splits and leakage-prevention measures is necessary to support the zero-shot and cross-dataset claims. The original manuscript summarized the overall training corpus but did not enumerate the exact partitioning. In the revised version we have inserted a new subsection in Methods that (i) details the train/validation/test allocation across the 30 datasets, (ii) confirms patient-level partitioning was used throughout, (iii) explicitly identifies the six datasets reserved for zero-shot evaluation, and (iv) describes the overlap checks performed to ensure no patient or study appeared in both training and held-out sets. These additions directly substantiate the reported generalization results. revision: yes

Circularity Check

0 steps flagged

No significant circularity in empirical performance claims

full rationale

The paper reports measured segmentation performance (mean DSC 0.87 on 47 targets, 0.85 on zero-shot sets) after decoder-only fine-tuning on 1.1M MRI slices, with direct comparisons to MedSAM and statistical testing. No equations, predictions, or first-principles derivations are present that reduce by construction to fitted inputs, self-citations, or ansatzes. The frozen-encoder choice is an explicit efficiency decision, not a load-bearing premise that loops back to the reported metrics. The analysis is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper relies on the standard SAM architecture and loss functions from prior work; no new physical or mathematical axioms are introduced. The main unstated premise is that natural-image pretraining transfers usefully to MRI when only the decoder is updated.

axioms (1)
  • domain assumption Freezing the SAM ViT-B/16 image encoder preserves representations that remain useful for MRI despite domain shift in contrast and intensity.
    Stated when the authors choose to freeze encoders to reduce compute and preserve pretrained features.

pith-pipeline@v0.9.0 · 5926 in / 1371 out tokens · 26146 ms · 2026-05-21T20:39:42.176570+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

74 extracted references · 74 canonical work pages · 4 internal anchors

  1. [1]

    A reliable framework for accurate brain image examination and treatment planning based on early diagnosis support for clinicians,

    S. L. Fernandes, U. J. Tanik, V. Rajinikanth, and K. A. Karthik, "A reliable framework for accurate brain image examination and treatment planning based on early diagnosis support for clinicians," Neural Computing and Applications, vol. 32, no. 20, pp. 15897-15908, 2020/10/01 2020, doi: 10.1007/s00521-019-04369-5

    work page doi:10.1007/s00521-019-04369-5 2020

  2. [2]

    Inter -observer variability of manual contour delineation of structures in CT,

    L. Joskowicz, D. Cohen, N. Caplan, and J. Sosna, "Inter -observer variability of manual contour delineation of structures in CT," European Radiology, vol. 29, no. 3, pp. 1391 -1399, 2019/03/01 2019, doi: 10.1007/s00330 - 018-5695-5

    work page doi:10.1007/s00330 2019

  3. [3]

    Magnetization transfer MRI in multiple sclerosis,

    M. Filippi and F. Agosta, "Magnetization transfer MRI in multiple sclerosis," Journal of Neuroimaging, vol. 17, pp. 22S-26S, 2007

    work page 2007

  4. [4]

    Nonrigid Registration of Joint Histograms for Intensity Standardization in Magnetic Resonance Imaging,

    F. Jager and J. Hornegger, "Nonrigid Registration of Joint Histograms for Intensity Standardization in Magnetic Resonance Imaging," IEEE Transactions on Medical Imaging, vol. 28, no. 1, pp. 137 -150, 2009, doi: 10.1109/TMI.2008.2004429

    work page doi:10.1109/tmi.2008.2004429 2009

  5. [5]

    Ultrahigh-field MRI: where it really makes a difference,

    S. Trattnig, G. Hangel, S. D. Robinson, V. Juras, P. Szomolanyi, and A. Dal-Bianco, "Ultrahigh-field MRI: where it really makes a difference," Die Radiologie, vol. 64, no. 1, pp. 1-8, 2024/11/01 2024, doi: 10.1007/s00117-023- 01184-x

    work page doi:10.1007/s00117-023- 2024

  6. [6]

    Comparison of Routine Brain Imaging at 3 T and 7 T,

    E. Springer et al., "Comparison of Routine Brain Imaging at 3 T and 7 T," Investigative Radiology, vol. 51, no. 8,

    work page

  7. [7]

    Available: https://journals.lww.com/investigativeradiology/fulltext/2016/08000/comparison_of_routine_brain_imaging _at_3_t_and_7_t.1.aspx

    [Online]. Available: https://journals.lww.com/investigativeradiology/fulltext/2016/08000/comparison_of_routine_brain_imaging _at_3_t_and_7_t.1.aspx

    work page 2016

  8. [8]

    Foundation models for biomedical image segmentation: A survey,

    H. H. Lee et al. , "Foundation models for biomedical image segmentation: A survey," arXiv preprint arXiv:2401.07654, 2024

    work page arXiv 2024

  9. [9]

    MRI at 7 Tesla and above: demonstrated and potential capabilities,

    O. Kraff, A. Fischer, A. M. Nagel, C. Mönninghoff, and M. E. Ladd, "MRI at 7 Tesla and above: demonstrated and potential capabilities," Journal of Magnetic Resonance Imaging, vol. 41, no. 1, pp. 13-33, 2015

    work page 2015

  10. [10]

    Quantitative susceptibility mapping: basic methods and clinical applications,

    T. Harada et al., "Quantitative susceptibility mapping: basic methods and clinical applications," Radiographics, vol. 42, no. 4, pp. 1161-1176, 2022

    work page 2022

  11. [11]

    STrategically Acquired Gradient Echo (STAGE) imaging, part III: Technical advances and clinical applications of a rapid multi -contrast multi-parametric brain imaging method,

    E. M. Haacke et al., "STrategically Acquired Gradient Echo (STAGE) imaging, part III: Technical advances and clinical applications of a rapid multi -contrast multi-parametric brain imaging method," Magnetic Resonance Imaging, vol. 65, pp. 15-26, 2020/01/01/ 2020, doi: https://doi.org/10.1016/j.mri.2019.09.006

    work page doi:10.1016/j.mri.2019.09.006 2020

  12. [12]

    Application of a Population Based Electrophysiological Database to the Planning and Guidance of Deep Brain Stereotactic Neurosurgery,

    K. W. Finnis, Y. P. Starreveld, A. G. Parrent, A. F. Sadikot, and T. M. Peters, "Application of a Population Based Electrophysiological Database to the Planning and Guidance of Deep Brain Stereotactic Neurosurgery," in Medical Image Computing and Computer-Assisted Intervention — MICCAI 2002, Berlin, Heidelberg, T. Dohi and R. Kikinis, Eds., 2002// 2002: S...

    work page 2002

  13. [13]

    A Multi -Atlas- Based Segmentation Framework for Prostate Brachytherapy,

    S. Nouranian, S. S. Mahdavi, I. Spadinger, W. J. Morris, S. E. Salcudean, and P. Abolmaesumi, "A Multi -Atlas- Based Segmentation Framework for Prostate Brachytherapy," IEEE Transactions on Medical Imaging, vol. 34, no. 4, pp. 950-961, 2015, doi: 10.1109/TMI.2014.2371823

    work page doi:10.1109/tmi.2014.2371823 2015

  14. [14]

    Automatic Substitute Computed Tomography Generation and Contouring for Magnetic Resonance Imaging (MRI) -Alone External Beam Radiation Therapy From Standard MRI Sequences,

    J. A. Dowling et al., "Automatic Substitute Computed Tomography Generation and Contouring for Magnetic Resonance Imaging (MRI) -Alone External Beam Radiation Therapy From Standard MRI Sequences," International Journal of Radiation Oncology*Biology*Physics, vol. 93, no. 5, pp. 1144-1153, 2015/12/01/ 2015, doi: https://doi.org/10.1016/j.ijrobp.2015.08.045

    work page doi:10.1016/j.ijrobp.2015.08.045 2015

  15. [15]

    Patient Specific Prostate Segmentation in 3 -D Magnetic Resonance Images,

    S. S. Chandra et al. , "Patient Specific Prostate Segmentation in 3 -D Magnetic Resonance Images," IEEE Transactions on Medical Imaging, vol. 31, no. 10, pp. 1955-1964, 2012, doi: 10.1109/TMI.2012.2211377. ` 17

    work page doi:10.1109/tmi.2012.2211377 1955

  16. [16]

    CNN: A Vision of Complexity,

    L. O. Chua, "CNN: A Vision of Complexity," International Journal of Bifurcation and Chaos, vol. 07, no. 10, pp. 2219-2425, 1997, doi: 10.1142/s0218127497001618

    work page doi:10.1142/s0218127497001618 1997

  17. [17]

    Medical Image Segmentation Review: The Success of U -Net,

    R. Azad et al. , "Medical Image Segmentation Review: The Success of U -Net," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 10076-10095, 2024, doi: 10.1109/TPAMI.2024.3435571

    work page doi:10.1109/tpami.2024.3435571 2024

  18. [18]

    3D U -Net: learning dense volumetric segmentation from sparse annotation,

    Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, "3D U -Net: learning dense volumetric segmentation from sparse annotation," in International conference on medical image computing and computer-assisted intervention, 2016: Springer, pp. 424-432

    work page 2016

  19. [19]

    Naushad and I

    Y. Zhu, A. Tchetchenian, X. Yin, A. Liew, Y. Song, and E. Meijering, "AC -UNet: Adaptive Connection UNet for White Matter Tract Segmentation Through Neural Architecture Search," in 2024 IEEE International Symposium on Biomedical Imaging (ISBI), 27-30 May 2024 2024, pp. 1-5, doi: 10.1109/ISBI56570.2024.10635404

    work page doi:10.1109/isbi56570.2024.10635404 2024

  20. [20]

    Architecture Analysis and Benchmarking of 3D U- Shaped Deep Learning Models for Thoracic Anatomical Segmentation,

    A. Harirpoush, A. Rasoulian, M. Kersten-Oertel, and Y. Xiao, "Architecture Analysis and Benchmarking of 3D U- Shaped Deep Learning Models for Thoracic Anatomical Segmentation," IEEE Access, vol. 12, pp. 127592-127603, 2024, doi: 10.1109/ACCESS.2024.3456674

    work page doi:10.1109/access.2024.3456674 2024

  21. [21]

    Attention U-Net: Learning Where to Look for the Pancreas

    O. Oktay et al., "Attention u -net: Learning where to look for the pancreas," arXiv preprint arXiv:1804.03999, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  22. [22]

    CAN3D: Fast 3D medical image segmentation via compact context aggregation,

    W. Dai et al., "CAN3D: Fast 3D medical image segmentation via compact context aggregation," Medical Image Analysis, vol. 82, p. 102562, 2022

    work page 2022

  23. [23]

    Mdu-net: Multi-scale densely connected u-net for biomedical image segmentation,

    J. Zhang, Y. Zhang, Y. Jin, J. Xu, and X. Xu, "Mdu-net: Multi-scale densely connected u-net for biomedical image segmentation," Health Information Science and Systems, vol. 11, no. 1, p. 13, 2023

    work page 2023

  24. [24]

    nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation

    F. Isensee et al. , "nnu -net: Self -adapting framework for u -net-based medical image segmentation," arXiv preprint arXiv:1809.10486, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  25. [25]

    Recurrent residual U -Net for medical image segmentation,

    M. Z. Alom, C. Yakopcic, M. Hasan, T. M. Taha, and V. K. Asari, "Recurrent residual U -Net for medical image segmentation," Journal of medical imaging, vol. 6, no. 1, pp. 014006-014006, 2019

    work page 2019

  26. [26]

    V -net: Fully convolutional neural networks for volumetric medical image segmentation,

    F. Milletari, N. Navab, and S. -A. Ahmadi, "V -net: Fully convolutional neural networks for volumetric medical image segmentation," in 2016 fourth international conference on 3D vision (3DV), 2016: Ieee, pp. 565-571

    work page 2016

  27. [27]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, "U-Net: Convolutional Networks for Biomedical Image Segmentation," Cham, 2015: Springer International Publishing, in Medical Image Computing and Computer -Assisted Intervention – MICCAI 2015, pp. 234-241

    work page 2015

  28. [28]

    Cellpose: a generalist algorithm for cellular segmentation,

    C. Stringer, T. Wang, M. Michaelos, and M. Pachitariu, "Cellpose: a generalist algorithm for cellular segmentation," Nature Methods, vol. 18, no. 1, pp. 100 -106, 2021/01/01 2021, doi: 10.1038/s41592 -020- 01018-x

    work page doi:10.1038/s41592 2021

  29. [29]

    Effects of MRI scanner manufacturers in classification tasks with deep learning models,

    R. Kushol, P. Parnianpour, A. H. Wilman, S. Kalra, and Y. -H. Yang, "Effects of MRI scanner manufacturers in classification tasks with deep learning models," Scientific Reports, vol. 13, no. 1, p. 16791, 2023/10/05 2023, doi: 10.1038/s41598-023-43715-5

    work page doi:10.1038/s41598-023-43715-5 2023

  30. [30]

    Performance changes due to differences among annotating radiologists for training data in computerized lesion detection,

    Y. Nomura et al., "Performance changes due to differences among annotating radiologists for training data in computerized lesion detection," International Journal of Computer Assisted Radiology and Surgery, vol. 19, no. 8, pp. 1527-1536, 2024/08/01 2024, doi: 10.1007/s11548-024-03136-9

    work page doi:10.1007/s11548-024-03136-9 2024

  31. [31]

    Slippage boosted spectral cleaning in a seeded free - electron laser ,

    F. Renard, S. Guedria, N. D. Palma, and N. Vuillerme, "Variability and reproducibility in deep learning for medical image segmentation," Scientific Reports, vol. 10, no. 1, p. 13724, 2020/08/13 2020, doi: 10.1038/s41598 -020- 69920-0

    work page doi:10.1038/s41598 2020

  32. [32]

    Visual Semantic Segmentation Based on Few/Zero -Shot Learning: An Overview,

    W. Ren, Y. Tang, Q. Sun, C. Zhao, and Q. L. Han, "Visual Semantic Segmentation Based on Few/Zero -Shot Learning: An Overview," IEEE/CAA Journal of Automatica Sinica, vol. 11, no. 5, pp. 1106 -1126, 2024, doi: 10.1109/JAS.2023.123207. ` 18

    work page doi:10.1109/jas.2023.123207 2024

  33. [33]

    From CNN to Transformer: A Review of Medical Image Segmentation Models,

    W. Yao, J. Bai, W. Liao, Y. Chen, M. Liu, and Y. Xie, "From CNN to Transformer: A Review of Medical Image Segmentation Models," Journal of Imaging Informatics in Medicine, vol. 37, no. 4, pp. 1529-1547, 2024/08/01 2024, doi: 10.1007/s10278-024-00981-7

    work page doi:10.1007/s10278-024-00981-7 2024

  34. [34]

    Segment anything,

    A. Kirillov et al., "Segment anything," in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015-4026

    work page 2023

  35. [35]

    Segment Anything Model Can Not Segment Anything: Assessing AI Foundation Model’s Generalizability in Permafrost Mapping,

    W. Li et al. , "Segment Anything Model Can Not Segment Anything: Assessing AI Foundation Model’s Generalizability in Permafrost Mapping," Remote Sensing, vol. 16, no. 5, p. 797, 2024. [Online]. Available: https://www.mdpi.com/2072-4292/16/5/797

    work page 2024

  36. [36]

    Segment anything model for medical image analysis: an experimental study,

    M. A. Mazurowski, H. Dong, H. Gu, J. Yang, N. Konz, and Y. Zhang, "Segment anything model for medical image analysis: an experimental study," Medical Image Analysis, vol. 89, p. 102918, 2023

    work page 2023

  37. [37]

    Segment anything model for medical images?,

    Y. Huang et al., "Segment anything model for medical images?," Medical Image Analysis, vol. 92, p. 103061, 2024/02/01/ 2024, doi: https://doi.org/10.1016/j.media.2023.103061

    work page doi:10.1016/j.media.2023.103061 2024

  38. [38]

    Boosting object proposals: From pascal to coco,

    J. Pont-Tuset and L. Van Gool, "Boosting object proposals: From pascal to coco," in Proceedings of the IEEE international conference on computer vision, 2015, pp. 1546-1554

    work page 2015

  39. [39]

    Magnetic resonance elastography from fundamental soft -tissue mechanics to diagnostic imaging,

    I. Sack, "Magnetic resonance elastography from fundamental soft -tissue mechanics to diagnostic imaging," Nature Reviews Physics, vol. 5, no. 1, pp. 25-42, 2023/01/01 2023, doi: 10.1038/s42254-022-00543-2

    work page doi:10.1038/s42254-022-00543-2 2023

  40. [40]

    Nature Communications15(1), 654 (1 2024)

    J. Ma, Y. He, F. Li, L. Han, C. You, and B. Wang, "Segment anything in medical images," Nature Communications, vol. 15, no. 1, p. 654, 2024/01/22 2024, doi: 10.1038/s41467-024-44824-z

    work page doi:10.1038/s41467-024-44824-z 2024

  41. [41]

    An efficient segment anything model for the segmentation of medical images,

    G. Dong et al. , "An efficient segment anything model for the segmentation of medical images," Scientific Reports, vol. 14, no. 1, p. 19425, 2024

    work page 2024

  42. [42]

    Ma -sam: Modality -agnostic sam adaptation for 3d medical image segmentation,

    C. Chen et al. , "Ma -sam: Modality -agnostic sam adaptation for 3d medical image segmentation," Medical Image Analysis, vol. 98, p. 103310, 2024

    work page 2024

  43. [43]

    Medical SAM adapter: Adapting segment anything model for medical image segmentation,

    J. Wu et al. , "Medical SAM adapter: Adapting segment anything model for medical image segmentation," Medical Image Analysis, vol. 102, p. 103547, 2025/05/01/ 2025, doi: https://doi.org/10.1016/j.media.2025.103547

    work page doi:10.1016/j.media.2025.103547 2025

  44. [44]

    Desam: Decoupled segment anything model for generalizable medical image segmentation,

    Y. Gao, W. Xia, D. Hu, W. Wang, and X. Gao, "Desam: Decoupled segment anything model for generalizable medical image segmentation," in International Conference on Medical Image Computing and Computer - Assisted Intervention, 2024: Springer, pp. 509-519

    work page 2024

  45. [45]

    MedLSAM: Localize and segment anything model for 3D CT images,

    W. Lei, W. Xu, K. Li, X. Zhang, and S. Zhang, "MedLSAM: Localize and segment anything model for 3D CT images," Medical Image Analysis, vol. 99, p. 103370, 2025

    work page 2025

  46. [46]

    Focal loss for dense object detection,

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, "Focal loss for dense object detection," in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2980-2988

    work page 2017

  47. [47]

    Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations,

    C. H. Sudre, W. Li, T. Vercauteren, S. Ourselin, and M. Jorge Cardoso, "Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations," in Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML -CDS 2017, Hel...

    work page 2017

  48. [48]

    Measures of the amount of ecologic association between species,

    L. R. Dice, "Measures of the amount of ecologic association between species," Ecology, vol. 26, no. 3, pp. 297- 302, 1945

    work page 1945

  49. [49]

    Deep learning techniques for automatic MRI cardiac multi -structures segmentation and diagnosis: is the problem solved?,

    O. Bernard et al. , "Deep learning techniques for automatic MRI cardiac multi -structures segmentation and diagnosis: is the problem solved?," IEEE transactions on medical imaging, vol. 37, no. 11, pp. 2514-2525, 2018

    work page 2018

  50. [50]

    Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation,

    Y. Ji et al., "Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation," Advances in neural information processing systems, vol. 35, pp. 36722-36732, 2022. ` 19

    work page 2022

  51. [51]

    Enhanced performance of brain tumor classification via tumor region augmentation and partition,

    J. Cheng et al. , "Enhanced performance of brain tumor classification via tumor region augmentation and partition," PloS one, vol. 10, no. 10, p. e0140381, 2015

    work page 2015

  52. [52]

    Retrieval of brain tumors by adaptive spatial pooling and fisher vector representation,

    J. Cheng et al., "Retrieval of brain tumors by adaptive spatial pooling and fisher vector representation," PloS one, vol. 11, no. 6, p. e0157112, 2016

    work page 2016

  53. [53]

    Segmentation labels and radiomic features for the pre -operative scans of the TCGA -GBM collection (2017),

    S. Bakas et al. , "Segmentation labels and radiomic features for the pre -operative scans of the TCGA -GBM collection (2017)," DOI: https://doi. org/10.7937 K, vol. 9

    work page 2017

  54. [54]

    Segmentation labels and radiomic features for the pre -operative scans of the TCGA -LGG collection [Data Set]. The Cancer Imaging Archive,

    S. Bakas et al. , "Segmentation labels and radiomic features for the pre -operative scans of the TCGA -LGG collection [Data Set]. The Cancer Imaging Archive," ed: Version, 2017

    work page 2017

  55. [55]

    Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features,

    S. Bakas et al., "Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features," Scientific data, vol. 4, no. 1, pp. 1-13, 2017

    work page 2017

  56. [56]

    Identifying the Best Machine Learning Algorithms for Brain Tumor Segmentation, Progression Assessment, and Overall Survival Prediction in the BRATS Challenge

    S. Bakas et al., "Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the BRATS challenge," arXiv preprint arXiv:1811.02629, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  57. [57]

    The multimodal brain tumor image segmentation benchmark (BRATS),

    B. H. Menze et al., "The multimodal brain tumor image segmentation benchmark (BRATS)," IEEE transactions on medical imaging, vol. 34, no. 10, pp. 1993-2024, 2014

    work page 1993

  58. [58]

    Tumor radiomic heterogeneity: Multiparametric functional imaging to characterize variability and predict response following cervical cancer radiation therapy,

    S. R. Bowen et al. , "Tumor radiomic heterogeneity: Multiparametric functional imaging to characterize variability and predict response following cervical cancer radiation therapy," Journal of Magnetic Resonance Imaging, vol. 47, no. 5, pp. 1388-1396, 2018

    work page 2018

  59. [59]

    CHAOS challenge -combined (CT -MR) healthy abdominal organ segmentation,

    A. E. Kavur et al. , "CHAOS challenge -combined (CT -MR) healthy abdominal organ segmentation," Medical image analysis, vol. 69, p. 101950, 2021

    work page 2021

  60. [60]

    CrossMoDA 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation,

    R. Dorent et al., "CrossMoDA 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation," Medical Image Analysis, vol. 83, p. 102628, 2023

    work page 2021

  61. [61]

    HVSMR-2.0: A 3D cardiovascular MR dataset for whole-heart segmentation in congenital heart disease,

    D. F. Pace et al., "HVSMR-2.0: A 3D cardiovascular MR dataset for whole-heart segmentation in congenital heart disease," Scientific Data, vol. 11, no. 1, p. 721, 2024/07/02 2024, doi: 10.1038/s41597-024-03469-9

    work page doi:10.1038/s41597-024-03469-9 2024

  62. [62]

    ISLES 2022: A multi -center magnetic resonance imaging stroke lesion segmentation dataset,

    M. R. Hernandez Petzsche et al. , "ISLES 2022: A multi -center magnetic resonance imaging stroke lesion segmentation dataset," Scientific data, vol. 9, no. 1, p. 762, 2022

    work page 2022

  63. [63]

    The Medical Segmentation Decathlon,

    M. Antonelli et al., "The Medical Segmentation Decathlon," Nature Communications, vol. 13, no. 1, p. 4128, 2022/07/15 2022, doi: 10.1038/s41467-022-30695-9

    work page doi:10.1038/s41467-022-30695-9 2022

  64. [64]

    A large annotated medical image dataset for the development and evaluation of segmentation algorithms

    A. L. Simpson et al. , "A large annotated medical image dataset for the development and evaluation of segmentation algorithms," arXiv preprint arXiv:1902.09063, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  65. [65]

    The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository,

    K. Clark et al., "The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository," Journal of Digital Imaging, vol. 26, no. 6, pp. 1045-1057, 2013/12/01 2013, doi: 10.1007/s10278-013-9622-7

    work page doi:10.1007/s10278-013-9622-7 2013

  66. [66]

    Evaluation of prostate segmentation algorithms for MRI: the PROMISE12 challenge,

    G. Litjens et al., "Evaluation of prostate segmentation algorithms for MRI: the PROMISE12 challenge," Medical image analysis, vol. 18, no. 2, pp. 359-373, 2014

    work page 2014

  67. [67]

    End -to-end prostate cancer detection in bpMRI via 3D CNNs: effects of attention mechanisms, clinical priori and decoupled false positive reduction,

    A. Saha, M. Hosseinzadeh, and H. Huisman, "End -to-end prostate cancer detection in bpMRI via 3D CNNs: effects of attention mechanisms, clinical priori and decoupled false positive reduction," Medical image analysis, vol. 73, p. 102155, 2021

    work page 2021

  68. [68]

    An annotated test-retest collection of prostate multiparametric MRI,

    A. Fedorov et al., "An annotated test-retest collection of prostate multiparametric MRI," Scientific data, vol. 5, no. 1, pp. 1-13, 2018

    work page 2018

  69. [69]

    Variability of manual segmentation of the prostate in axial T2-weighted MRI: a multi-reader study,

    A. S. Becker et al., "Variability of manual segmentation of the prostate in axial T2-weighted MRI: a multi-reader study," European journal of radiology, vol. 121, p. 108716, 2019

    work page 2019

  70. [70]

    Robust detection and segmentation for diagnosis of vertebral diseases using routine MR images,

    D. Zukić, A. Vlasák, J. Egger, D. Hořínek, C. Nimsky, and A. Kolb, "Robust detection and segmentation for diagnosis of vertebral diseases using routine MR images," in Computer Graphics Forum , 2014, vol. 33, no. 6: Wiley Online Library, pp. 190-204. ` 20

    work page 2014

  71. [71]

    Totalsegmentator mri: Robust sequence -independent segmentation of multiple anatomic structures in mri,

    T. A. D'Antonoli et al. , "Totalsegmentator mri: Robust sequence -independent segmentation of multiple anatomic structures in mri," Radiology, vol. 314, no. 2, p. e241613, 2025

    work page 2025

  72. [72]

    Standardized assessment of automatic segmentation of white matter hyperintensities and results of the WMH segmentation challenge,

    H. J. Kuijf et al., "Standardized assessment of automatic segmentation of white matter hyperintensities and results of the WMH segmentation challenge," IEEE transactions on medical imaging, vol. 38, no. 11, pp. 2556- 2568, 2019

    work page 2019

  73. [73]

    Automated segmentation of knee bone and cartilage combining statistical shape knowledge and convolutional neural networks: Data from the Osteoarthritis Initiative,

    F. Ambellan, A. Tack, M. Ehlke, and S. Zachow, "Automated segmentation of knee bone and cartilage combining statistical shape knowledge and convolutional neural networks: Data from the Osteoarthritis Initiative," Medical image analysis, vol. 52, pp. 109-118, 2019

    work page 2019

  74. [74]

    Medical image analysis using improved SAM -Med2D: segmentation and classification perspectives,

    J. Sun et al. , "Medical image analysis using improved SAM -Med2D: segmentation and classification perspectives," BMC Medical Imaging, vol. 24, no. 1, p. 241, 2024. ` 1 Supplementary Table S1: 36 MRI datasets Dataset Name Modality Segmentation Targets # of samples Labeled_Slices Num_of_pairs ACDC [48] *# MR Heart anatomies 300 2839 7991 AMOS_MR [49] # MR ...

    work page 2024