Pith. sign in

REVIEW 4 major objections 4 minor 35 references

RFMedSAM 2: Automatic Prompt Refinement for Enhanced Volumetric Medical Image Segmentation with SAM 2

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RFMedSAM 2 pairs a U-Net prompt generator with two SAM 2 refinement stages and claims Dice scores of 90.7% on AMOS2022 and 86.7% on BTCV, surpassing nnUNet baselines.

desk verdict Solid engineering paper with a load-bearing baseline problem: the SOTA margins rely on nnUNet numbers the authors did not reproduce, so the claims need a controlled re-run before they can be trusted. read the letter →

arxiv 2502.02741 v1 pith:IIIOSRON submitted 2025-02-04 cs.CV

Bin Xie , Hao Tang , Yan Yan , Gady Agam This is my paper · ORCID
classification cs.CV
keywords SAM2medicalimagesegmentationpromptrefinementadapterfine-tuningU-NetgeneratorDicescoreCT
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that the Segment Anything Model 2 (SAM 2), a prompt-driven foundation model, can be turned into a fully automatic, state-of-the-art volumetric medical image segmenter. The authors first measure SAM 2's upper limit: after adding lightweight adapters to its image encoder, mask decoder, memory encoder, and memory attention, and fixing a temporal-positioning flaw in its memory attention, SAM 2 with ground-truth bounding boxes reaches a Dice score of 92.3% on the BTCV dataset. They then remove the need for manual prompts by adding an independent U-Net that predicts masks and bounding boxes; those boxes are fed to SAM 2, whose two refinement stages (Step 1 and Step 2) progressively fix the U-Net's errors. The final system reports 90.7% Dice on AMOS2022 and 86.7% on BTCV, surpassing the nnUNet baselines cited in the paper by 2.9 and 6.4 points respectively. A reader would care because this is evidence that a generic foundation model, with automatic prompt generation and staged refinement, can compete with hand-crafted medical segmentation pipelines.

What carries the argument

The load-bearing pieces are three. (1) DWConvAdapter: a residual adapter block that adds depth-wise convolution, layer normalization, and GeLU after the self/cross-attention blocks that operate on image embeddings, inserted into Hiera (the image encoder), memory attention, and mask decoder; it captures spatial context while retaining a parallel skip so the original features survive even if the adapter learns nothing. (2) CNN-Adapter: a point-wise-downsample, depth-wise-convolve, point-wise-upsample bottleneck for SAM 2's convolutional layers, again with a skip connection. (3) The corrected memory-attention schedule: instead of giving all prompted frames temporal position 0, the authors set the current frame as position 0 and attend to up to six preceding frames, which removes false-positive attention artifacts and gives the Step-2 refinement usable temporal context. The automatic-prompt branch uses a 3D U-Net (with skip connections and deep supervision) whose predicted masks are converted to bounding boxes for Step 1, and the masks predicted in Step 1 are converted to new boxes for Step 2. The claimed mechanism is that each refinement re-derives boxes from its own partial masks, converging toward the 92.3% prompt-bound performance.

What would settle it

Run nnUNet under the exact 5-fold cross-validation splits and preprocessing used for RFMedSAM 2 on AMOS2022 and BTCV, compute Dice with the same implementation, and compare; if nnUNet equals or exceeds 90.7% and 86.7%, the paper's margin claims collapse.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is twofold. First, SAM 2's ceiling with optimal prompts is high: fine-tuning only with small adapters — a depth-wise convolutional adapter for attention blocks that process image embeddings and a CNN adapter for convolutional layers — plus a corrected memory-attention frame selection (making the current frame the temporal position 0 and attending to up to six previous frames) yields 92.30% Dice on BTCV, about 12 percentage points above the paper's nnUNet baseline. Second, this prompt-dependent model can be made autonomous without losing the edge: a separately trained U-Net generates initial masks, from which bounding boxes are derived; SAM 2's two refinement stages use those boxes as prompts, then re-derive boxes from its own predicted masks for a second refinement. The reported results are 90.7% on AMOS2022 and 86.7% on BTCV, both above the paper's nnUNet comparisons. The authors interpret this as showing that the remaining gap between SAM 2 with perfect prompts and SAM 2 with generated prompts is modest, and that the prompt generator's design matters: learning masks before boxes and keeping the generator separate from SAM 2's image encoder avoids conflicting training losses.

Load-bearing premise

The claim that RFMedSAM 2 beats nnUNet depends on the cited nnUNet scores (87.8% on AMOS, 80.2% on BTCV) being produced under the same data splits, preprocessing, and evaluation protocol as RFMedSAM 2; the paper does not re-run nnUNet in its own pipeline, and the BTCV baseline is unusually low.

Editorial extensions

If this is right

  • If the reported margins hold, a single SAM 2 checkpoint with adapter layers and a small U-Net could replace task-specific nnUNet pipelines for multi-organ abdominal CT segmentation.
  • The two-stage refinement means the system's accuracy is not capped by the prompt generator's own Dice; each Step 1 and Step 2 mask can fix regions the U-Net missed.
  • The 92.3% result with ground-truth boxes quantifies how much headroom remains between prompt-free and perfectly-prompted operation.
  • Keeping the prompt generator as an independent U-Net avoids the training instability the authors observed when the image encoder received both auxiliary mask losses and SAM 2 output losses.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the protocol-comparison caveat is resolved, this architecture would be a natural testbed for other modalities (MRI, ultrasound) and for video volumes, where SAM 2's memory attention should help temporal consistency; the paper does not test those.
  • The observation that object-score learning hurts the learnable-mask prompt generator but helps learnable boxes suggests a design heuristic: let masks carry the presence/absence signal rather than a separate scalar head.
  • The stepwise improvement (U-Net 0.856 -> Step 1 0.864 -> Step 2 0.867 on BTCV) implies that further refinement stages might yield diminishing returns, but each stage is cheap in parameters since only adapters are updated; this could be tested by adding a third refinement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes RFMedSAM 2, an adaptation of SAM 2 to volumetric medical image segmentation. The method inserts DWConvAdapters and CNN-Adapters into SAM 2's image encoder, mask decoder, memory encoder, and memory attention, modifies the temporal frame-selection strategy, and adds an independent U-Net that generates initial masks and bounding boxes as prompts. A dual-stage SAM 2 refinement then produces the final segmentation. On AMOS2022 the authors report a mean Dice of 90.7%, surpassing their cited nnUNet baseline by 2.9%, and on BTCV they report 86.7%, surpassing nnUNet by 6.4%; with ground-truth bounding-box prompts the modified SAM 2 reaches 92.3% on BTCV. The paper includes ablations on frame selection, adapter placement, patch size, UNet architecture, and prompt-generator design, and a supplementary analysis of why prompt generators coupled to the image encoder cause unstable training.

Significance. The engineering contribution is plausible and the ablation study is informative: the paper carefully diagnoses SAM 2's temporal-position assignment, shows incremental gains from Step 0/1/2 refinement, and documents a concrete failure mode (loss oscillation) when the prompt generator shares the image encoder. If the reported gains survive a controlled comparison, the method would be a practically useful recipe for adapting SAM 2 to multi-organ CT segmentation. The main significance-limiting factors are the absence of a controlled re-run of the principal baseline (nnUNet), the unclear evaluation protocol on AMOS, the very small BTCV validation set, and no code or trained-model release. The claims are internally consistent across tables, but the headline state-of-the-art assertion rests on baseline numbers that the paper did not reproduce.

major comments (4)
  1. [§4.2.2, Table 3] The headline 6.4% improvement over nnUNet on BTCV is computed against a cited nnUNet Dice of 0.802, which is markedly below commonly reported nnUNet results on this dataset and is also below the 0.856 Dice of the authors' own Step-0 U-Net in Table 5 on the same validation split. Because nnUNet was not re-run under the authors' 24/4 split, preprocessing, and evaluation protocol, the margin may substantially reflect baseline underperformance. A controlled re-run of nnUNet under the identical pipeline, with per-case results, is required before this SOTA claim can be accepted.
  2. [§4.1 vs. Table 2 caption] The text states that AMOS evaluation uses the AMOS22 leaderboard, while the caption of Table 2 states that all results are based on 5-fold cross-validation without ensemble techniques. These protocols are incompatible: leaderboard evaluation uses a held-out test set, whereas 5-fold cross-validation produces internal test-fold predictions. It is therefore unclear whether the 0.907 RFMedSAM 2 result and the 0.878 nnUNet result come from the same test set. The authors must specify which numbers are official leaderboard submissions and which are internal 5-fold CV, and ensure all comparators are evaluated on the same protocol.
  3. [Table 3, §4.3] All BTCV conclusions are based on averages over only 4 validation cases, and no standard deviations, per-case Dice values, or significance tests are reported. Several per-organ results actually decrease relative to the cited nnUNet row (e.g., gallbladder 0.611 vs. 0.704), so the 6.4% overall margin is fragile. Per-case breakdowns and error bars, or a larger validation set, are needed to support the claimed state-of-the-art performance.
  4. [Section 11, supplementary] The implementation details specify hyperparameters for the U-Net (initial learning rate 0.001, SGD momentum 0.99, weight decay 3e-5, 1000 epochs, 250 iterations per epoch, deep supervision weights) but not for the SAM 2 fine-tuning stages: adapter learning rates, number of trainable vs. frozen parameters, training epochs for each refinement stage, and the loss weighting between the U-Net prompt generator and the SAM 2 outputs. Without these details, and without code release, the experiments cannot be reproduced from the manuscript alone.
minor comments (4)
  1. [Abstract vs. Section 1] The reported AMOS improvement over nnUNet is 2.9% in the Abstract and Section 4.2.1, but the contribution bullet in Section 1 states 2.7%; these numbers should be reconciled.
  2. [Section 9.2] The heading 'Modified Mask Encoder' describes modifications to the mask decoder, not the mask encoder; the heading and the text should be aligned.
  3. [Table 3] The table lists 'VNet [20]' as a convolution-based method, but reference [20] is Ronneberger et al.'s U-Net paper; the VNet citation appears to be incorrect.
  4. [Section 11] There is a typo 'weighted decay' (should be 'weight decay'), and the sentence 'we employ the deep supervision loss' should begin with a capital letter; in Eq. (1), the symbol e is used without an explicit definition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RFMedSAM 2's claims are empirical measurements on external test sets, and the paper's self-citations are confined to related work and are not load-bearing.

full rationale

The paper's central claims are benchmark results (DSC on AMOS2022 and BTCV), obtained by training and evaluating a concrete segmentation pipeline; no result is derived from an equation whose inputs already contain the output. The GT-prompt upper-bound experiment (92.30% on BTCV) is a standard evaluation where ground-truth boxes are inputs, not predictions, so it is not circular. The ablation tables (Tables 4-7) report measured Dice scores for alternative designs; none of these scores is a fitted parameter renamed as a prediction. The only self-citations, MaskSAM [24] and Self-Prompt SAM [25], appear in the related-work survey and are not used to justify an architectural choice, a uniqueness claim, or a baseline number, so they are not load-bearing. The remaining concerns, such as the cited nnUNet baselines not being re-run under the authors' protocol and the inconsistency between AMOS leaderboard evaluation and Table 2's 5-fold CV statement, are experimental-comparability or correctness risks rather than circularity; they do not make the derivation self-referential. Accordingly, no circular step meeting the required quotation-and-reduction standard is present.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central empirical claims rest on standard supervised learning assumptions plus a set of training hyperparameters and design choices tuned on the validation folds of the two datasets. There are no newly postulated physical or mathematical entities. The main non-standard premise is the comparability of external baseline numbers.

free parameters (7)
  • Initial learning rate = 0.001
    Set by the authors; sensitivity not reported.
  • Optimizer momentum = 0.99
    Chosen for SGD; no ablation.
  • Weight decay = 3e-5
    Chosen; no ablation.
  • Max epochs and iterations per epoch = 1000 epochs, 250 iterations per epoch
    Training budget; no ablation.
  • Input patch size = (32, 256, 256)
    Selected from three patch sizes in Table 6; higher DSC at larger depth.
  • Frame selection strategy = current frame as temporal position 0, up to 6 previous frames
    Selected from ablations in Figure 3(2); improves DSC from 90.74% to 91.58%.
  • Deep supervision weights = halving with resolution, normalized to sum 1
    Heuristic weighting; no sensitivity analysis.
assumptions (4)
  • domain assumption Pre-trained SAM 2 image features are transferable to medical imaging via adapters
    The entire fine-tuning approach assumes that Hiera features from natural images are a useful starting point for CT slices; demonstrated only empirically.
  • domain assumption Dice Similarity Coefficient averaged over organs is the appropriate comparison metric
    The paper reports only mean DSC; other metrics (surface distance, volume error) are not reported.
  • domain assumption Baseline numbers from prior publications are comparable to the reported results
    The improvements over nnUNet rely on tabulated numbers from previous papers, which may use different splits or protocols.
  • standard math Standard stochastic gradient descent with backpropagation optimizes the combined UNet and SAM 2 pipeline
    The training procedure in Section 11 assumes standard optimization works as intended.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RFMedSAM 2: Automatic Prompt Refinement for Enhanced Volumetric Medical Image Segmentation with SAM 2." pith.science (2026). https://pith.science/paper/IIIOSRON

@misc{pith2026250202741,
  author       = {Pith},
  title        = {Pith review of: RFMedSAM 2: Automatic Prompt Refinement for Enhanced Volumetric Medical Image Segmentation with SAM 2},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IIIOSRON}},
  note         = {Machine review of arXiv:2502.02741}
}
read the original abstract

Segment Anything Model 2 (SAM 2), a prompt-driven foundation model extending SAM to both image and video domains, has shown superior zero-shot performance compared to its predecessor. Building on SAM's success in medical image segmentation, SAM 2 presents significant potential for further advancement. However, similar to SAM, SAM 2 is limited by its output of binary masks, inability to infer semantic labels, and dependence on precise prompts for the target object area. Additionally, direct application of SAM and SAM 2 to medical image segmentation tasks yields suboptimal results. In this paper, we explore the upper performance limit of SAM 2 using custom fine-tuning adapters, achieving a Dice Similarity Coefficient (DSC) of 92.30% on the BTCV dataset, surpassing the state-of-the-art nnUNet by 12%. Following this, we address the prompt dependency by investigating various prompt generators. We introduce a UNet to autonomously generate predicted masks and bounding boxes, which serve as input to SAM 2. Subsequent dual-stage refinements by SAM 2 further enhance performance. Extensive experiments show that our method achieves state-of-the-art results on the AMOS2022 dataset, with a Dice improvement of 2.9% compared to nnUNet, and outperforms nnUNet by 6.4% on the BTCV dataset.

Figures

Figures reproduced from arXiv: 2502.02741 by the authors.

Figure 1
Figure 1. Overview of our proposed RFMedSAM 2. often suffer from limited high-quality annotations, which hampers the training of large-scale models. Consequently, architectures with higher inductive biases, such as CNNs, have been more easily trained from scratch to achieve strong performance in medical segmentation tasks. Foundation models [7, 10], trained on vast datasets, have shown remarkable capabilities in zero-shot and… view at source ↗
Figure 2
Figure 2. Overview of SAM 2. The pipeline includes steps for processing prompted and unprompted frames. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (1) Performance comparisons based on proposed methods. (2) Ablation studies for frame selection strategies. (3) Proposed [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on BTCV dataset. RFMedSAM 2 is the most precise for each class and has fewer segmentation outliers. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparisons with different output predictions for Step [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Benefits for refinement by Step 2. 82.77% to 81.17% Dice, we find a potential refinement ben￾efit illustrated in [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Details of the whole architecture of RFMedSAM 2. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: More visualization of two refinements. channels while reducing the spatial resolution to capture different deep-level features. The decoder upsamples the feature maps using transposed convolutions to restore spa￾tial resolution and refine predictions. Skip connections …
Figure 9
Figure 9. Figure 9: Oscillated losses if prompt generator built with image [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 20 canonical work pages

  1. [1]

    SAM3D: Segment Anything Model in Volumetric Medical Images

    Nhat-Tan Bui, Dinh-Hieu Hoang, Minh-Triet Tran, and Ngan Le. Sam3d: Segment anything model in volumetric medical images. arXiv preprint arXiv:2309.03493, 2023. 7

    work page Pith review arXiv 2023

  2. [2]

    Sam3d: Segment anything model in volumetric medical images

    Nhat-Tan Bui, Dinh-Hieu Hoang, Minh-Triet Tran, Gian- franco Doretto, Donald Adjeroh, Brijesh Patel, Arabinda Choudhary, and Ngan Le. Sam3d: Segment anything model in volumetric medical images. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE,

    2024

  3. [3]

    Swin-unet: Unet-like pure transformer for medical image segmentation

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xi- aopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 8

    arXiv 2021

  4. [4]

    Transunet: Transformers make strong encoders for medi- cal image segmentation

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medi- cal image segmentation. arXiv preprint arXiv:2102.04306,

    arXiv

  5. [5]

    Sam-med2d

    Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji- long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 7

    arXiv 2023

  6. [6]

    Sam-u: Multi-box prompts triggered uncertainty estimation for reliable sam in medical image

    Guoyao Deng, Ke Zou, Kai Ren, Meng Wang, Xuedong Yuan, Sancong Ying, and Huazhu Fu. Sam-u: Multi-box prompts triggered uncertainty estimation for reliable sam in medical image. In International Conference on Medical Im- age Computing and Computer-Assisted Intervention , pages 368–377. Springer, 2023. 1, 2, 7

    work page 2023

  7. [7]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2

    arXiv 2018

  8. [8]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images

    Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI Brainlesion Workshop, pages 272–284. Springer, 2021. 6, 7

    work page 2021

Show all 35 references
  1. [9]

    Unetr: Transformers for 3d medical image segmentation

    Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 574–5...

    2022

  2. [10]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16000– 16009, 2022. 1, 2

    2022

  3. [11]

    Automated design of deep learning methods for biomedical image segmentation

    Fabian Isensee, Paul F J ¨ager, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. Automated design of deep learning methods for biomedical image segmentation. arXiv preprint arXiv:1904.08128, 2019. 1, 2, 6, 7

    1904 arXiv

  4. [12]

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

    Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi- organ benchmark for versatile medical image segmentation. Advances in Neural Information Processing Systems , 35:...

    2022

  5. [13]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. arXiv preprint arXiv:2304.02643, 2023. 1, 2, 6, 7

    2023 arXiv

  6. [14]

    Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge

    B Landman, Z Xu, J Eugenio Igelsias, M Styner, T Langerak, and A Klein. Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge. In Proc. MICCAI: Multi- Atlas Labeling Beyond Cranial Vault-Workshop Challenge ,

  7. [15]

    3d ux-net: A large kernel volumetric convnet modernizing hierarchical transformer for medical image seg- mentation

    Ho Hin Lee, Shunxing Bao, Yuankai Huo, and Bennett A Landman. 3d ux-net: A large kernel volumetric convnet modernizing hierarchical transformer for medical image seg- mentation. arXiv preprint arXiv:2209.15076, 2022. 7

    2022 arXiv

  8. [16]

    Segment anything in medical images

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 2, 6, 7

    2024

  9. [17]

    Gpt-4 technical report

    R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2, 2023. 1

    2023 arXiv

  10. [18]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  11. [19]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 6, 7

    2024 arXiv

  12. [20]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In International Conference on Medical image com- puting and computer-assisted intervention , pages 234–241. Springer, 2015. 1, 2, 7

    2015

  13. [21]

    Hi- era: A hierarchical vision transformer without the bells-and- whistles

    Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hi- era: A hierarchical vision transformer without the bells-and- whistles. In International Conference on Machine Learn...

    2023

  14. [22]

    Autosam: Adapting sam to medical images by overloading the prompt encoder

    Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 2

    2023 arXiv

  15. [23]

    Transbts: Multimodal brain tumor seg- mentation using transformer

    Wenxuan Wang, Chen Chen, Meng Ding, Hong Yu, Sen Zha, and Jiangyun Li. Transbts: Multimodal brain tumor seg- mentation using transformer. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th In- ternational Conference, Strasbourg, France, September 2...

    2021

  16. [24]

    Masksam: Towards auto-prompt sam with mask classifi- cation for medical image segmentation

    Bin Xie, Hao Tang, Bin Duan, Dawen Cai, and Yan Yan. Masksam: Towards auto-prompt sam with mask classifi- cation for medical image segmentation. arXiv preprint arXiv:2403.14103, 2024. 1, 2

    2024 arXiv

  17. [25]

    Self-prompt sam: Medical image segmentation via auto- matic prompt sam adaptation, 2025

    Bin Xie, Hao Tang, Dawen Cai, Yan Yan, and Gady Agam. Self-prompt sam: Medical image segmentation via auto- matic prompt sam adaptation, 2025. 2

    2025

  18. [26]

    Customized segment any- thing model for medical image segmentation

    Kaidong Zhang and Dong Liu. Customized segment any- thing model for medical image segmentation. arXiv preprint arXiv:2304.13785, 2023. 1, 6, 7

    2023 arXiv

  19. [27]

    Segment any- thing model for medical image segmentation: Current ap- plications and future directions

    Yichi Zhang, Zhenrong Shen, and Rushi Jiao. Segment any- thing model for medical image segmentation: Current ap- plications and future directions. Computers in Biology and Medicine, page 108238, 2024. 1, 2

    2024

  20. [28]

    nnformer: Interleaved transformer for volumetric segmentation

    Hong-Yu Zhou, Jiansen Guo, Yinghao Zhang, Lequan Yu, Liansheng Wang, and Yizhou Yu. nnformer: Interleaved transformer for volumetric segmentation. arXiv preprint arXiv:2109.03201, 2021. 1, 2, 6, 7, 8 RFMedSAM 2: Automatic Prompt Refinement for Enhanced Volumetric Medical Image...

    2021 arXiv

  21. [29]

    Intrinsic issues of SAM2 Figure 2 illustrates the whole pipeline of SAM 2, highlight- ing several intrinsic issues for medical image segmentation. i) Omission to predict the first few frames: The first frame in two objects is the second frame, therefore, SAM 2 begins processin...

  22. [30]

    Potentials to force Step 2 for all frames. When we provide prompts at each frame for each class, SAM 2 does not process Step 2 and does not leverage the capabilities of Memory Attention, which can build relations with previous frames and prompted frames. To explore this functi...

  23. [31]

    Since the image encoder, the memory attention, and the mask decoder contain attention blocks for image embed- ding, which includes significant spatial information

    Motivation Behind the Designed Adapters. Since the image encoder, the memory attention, and the mask decoder contain attention blocks for image embed- ding, which includes significant spatial information. There- fore, we design the depth-wise convolutional adaption (DW- ConvAd...

  24. [32]

    Architecture of RFMedSAM 2 Figure 7(a) illustrates the overall pipeline and architecture of RFMedSAM 2, which consists of three primary steps. In Step 0, an additional UNet model is employed to take medical images as input, generating initial multi-class mask predictions, whic...

  25. [33]

    The features with a lower resolution gradually increase the resolution by convolution layers and then combined with higher resolution features

    Impact of Auxiliary Losses on Image En- coder Parameter Updates if Prompt Gen- erator Built with Image Encoder Figure 3(4d) illustrates a hierarchical structure with con- volutional layers combined with multi-level features from the image encoder. The features with a lower res...

  26. [34]

    We set the initial learning rate to 0.001 and employ a “poly” decay strategy in Eq

    Implementation Details We utilize some data augmentations such as rotation, scal- ing, Gaussian noise, Gaussian blur, brightness, and contrast adjustment, simulation of low resolution, gamma augmen- tation, and mirroring. We set the initial learning rate to 0.001 and employ a ...

  27. [35]

    With the two re- finements, the results clearly illustrate the progressive im- provement in segmentation accuracy, emphasizing the ef- fectiveness of our model’s refinement process

    More Visualization of Two Refinements In Figure 8, we present additional qualitative results show- casing the refinements at different stages. With the two re- finements, the results clearly illustrate the progressive im- provement in segmentation accuracy, emphasizing the ef-...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.