Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Fully Automated SAM for Single-source Domain Generalization in Medical Image Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that SAM can segment medical images from unseen clinical sites automatically, by generating its own bounding-box prompts from uncertainty-modeled features and by making the SAM mask decoder tolerant of imperfect prompts.

desk verdict A practical auto-prompting SAM pipeline, but target-domain leakage in model selection inflates the headline Dice gain; the fix is straightforward and the idea is worth refereeing. read the letter →

arxiv 2507.17281 v1 pith:ZK37E72Q submitted 2025-07-23 cs.CV

classification cs.CV
keywords single-sourcedomaingeneralizationSegmentAnythingModelmedicalimagesegmentationautomatedpromptgenerationfeatureuncertaintymodelingimage-promptembeddingfusionshiftprostateMRI
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 proposes FA-SAM, a framework that lets the Segment Anything Model (SAM) segment medical images from unseen clinical sites without any human-supplied prompts. The central claim is that a trainable Auto-prompted Generation Model (AGM) branch, equipped with Shallow Feature Uncertainty Modeling (SUFM), can produce the bounding-box prompts SAM needs on target domains, and that an Image-Prompt Embedding Fusion (IPEF) module makes SAM less misled when those boxes are imperfect. On prostate MRI, FA-SAM raises average Dice from 79.54% for the best prior SAM-based method to 84.53%; on fundus vessel datasets it reports the best average Dice as well. If valid, the approach would remove a major obstacle to deploying SAM in clinical settings: the reliance on expert-annotated prompts.

What carries the argument

The load-bearing object is the SUFM module inside the AGM branch. SUFM takes the channel-wise mean $\mu$ and standard deviation $\sigma$ of shallow encoder features, estimates their batch-level uncertainty, samples perturbed statistics from $\mathcal{N}(\mu, \Sigma_\mu^2)$ and $\mathcal{N}(\sigma, \Sigma_\sigma^2)$, adds Poisson-distributed noise, and renormalizes the features via $\hat{f} = \gamma(x)\cdot (x-\mu(x))\cdot\sigma(x)^{-1} + \beta(x)$. This simulates target-domain feature statistics from source data alone, so the AGM's predicted mask leads to a better bounding-box prompt for SAM. The second mechanism is IPEF, which concatenates multi-scale SAM image embeddings with prompt embeddings, passes them through SEResBlocks, and upsamples with skip connections before the mask decoder, letting SAM recover local detail even when the box is too small or too large. Together these two modules carry the claim that SAM can be fully automated and made robust to poor prompts.

What would settle it

For each target site, compute the actual channel-wise mean and variance of the AGM's shallow features and check whether they are drawn from the distribution SUFM samples; a site whose statistics fall far outside that envelope, or whose generated boxes consistently cut off the target structure, would disconfirm the paper's core claim. A direct test is to report box-IoU between AGM-generated boxes and ground-truth masks per site: box-IoU near chance would show the automated prompting does not work even if final Dice is aided by IPEF.

Watch

Extended reading notes

Core claim

FA-SAM claims fully automated SAM-based segmentation under single-source domain generalization. The method treats SAM's prompt dependency as the bottleneck: instead of asking an expert for bounding boxes, a separate segmentation network (AGM) predicts a preliminary mask on the target image, keeps the largest connected component via breadth-first search, and converts it into a box prompt. Domain shift is handled by SUFM, which models the channel-wise mean and variance of the AGM encoder's shallow features as uncertain, samples perturbed statistics from a multivariate Gaussian, and adds Poisson-distributed noise to simulate real medical noise. The IPEF module then concatenates multi-scale SAM image embeddings with prompt embeddings, enhances them with SEResBlocks, and upsamples with skip connections so the mask decoder recovers local detail that a too-small or too-large box would otherwise erase. The paper reports the best average Dice on both benchmark suites and a 4.99% average gain over SAMMed on prostate.

Load-bearing premise

Everything rests on the assumption that unseen target domains can be simulated by perturbing the channel-wise mean and variance of shallow features with Gaussian samples and Poisson noise; if actual target shifts involve new anatomy, field-of-view changes, or artifacts these perturbations do not capture, the AGM will emit poor boxes and the automated pipeline loses its advantage.

Editorial extensions

If this is right

  • SAM can be prompted without expert annotation: at test time the AGM branch generates the boxes, so FA-SAM runs end-to-end from raw target image to mask.
  • On the prostate cross-site benchmark, average Dice over six sites improves to 84.53%, a 4.99% gain over the best SAM-based comparison, with the largest per-site gains on sites B and C.
  • Fusing image and prompt embeddings with IPEF also helps when prompts are ground-truth boxes: SAM with IPEF and ground-truth prompts averages 91.54% versus 90.53% without IPEF.
  • SUFM works best when inserted between the first and second convolutional blocks and when Gaussian and Poisson disturbances are combined, according to the paper's ablations.

Reading between the lines

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

  • A straightforward extension would evaluate the generated bounding boxes themselves, measuring box IoU and size ratio relative to the mask on each target site; the paper reports final Dice but not box quality, so box quality is the unstated intermediate claim.
  • If the SUFM assumption holds, the same auto-prompting recipe could apply to other prompt-based vision foundation models, with uncertainty modeled on whatever encoder features they expose.
  • The reported gains leave open whether the method generalizes to 3D volumes, CT or ultrasound, or small lesions; testing those modalities would separate the benefit of automated prompting from the benefit of the two specific datasets used here.
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 / 6 minor

Summary. The paper proposes FA-SAM, a framework for fully automated SAM-based medical image segmentation under single-source domain generalization. It has two main components: an Auto-prompted Generation Model (AGM) branch with a Shallow Feature Uncertainty Modeling (SUFM) module that generates bounding-box prompts for unseen target domains, and an Image-Prompt Embedding Fusion (IPEF) module inserted in the SAM mask decoder to reduce sensitivity to poor prompts. The method is evaluated on a six-site prostate MRI dataset and a three-site fundus vessel dataset, reporting average Dice improvements over prior SAM-based and data-augmentation SDG baselines, including a claimed 4.99% average Dice improvement over SAMMed on prostate.

Significance. If the quantitative claims held, FA-SAM would address a genuine clinical bottleneck: removing the need for expert prompts while improving SAM's robustness under domain shift. The paper attacks a worthwhile problem, uses publicly available datasets, and includes ablations of the two proposed modules. The idea of coupling feature-statistics uncertainty with prompt generation is plausible and worth investigating. However, the current manuscript's evaluation protocol undermines the central generalization claim, because key design choices are selected using target-domain performance. The paper also lacks error bars and statistical tests, and some textual claims are contradicted by the tables. These issues must be resolved before the results can be interpreted as evidence of source-only domain generalization.

major comments (4)
  1. [Section IV-D (Tables IV and V)] The model-selection protocol leaks target-domain information into the reported generalization results. Table IV selects the SUFM insertion position '0-1' because it yields the best Final Prediction Dice on sites B–F, and Table V selects the combined Gaussian+Poisson distribution for the same reason. These choices are made using the very unseen domains on which the method is later evaluated in Table I, so the reported 84.53% average is not the performance of a fixed source-only model. The comparison with SAMMed and other baselines in Table I is therefore not a clean single-source generalization comparison. Please fix all hyperparameters and architectural choices using only the source domain (e.g., a source-domain validation split) and re-report all main results, or present the current numbers as an oracle-model-selection upper bound.
  2. [Tables I and II] All reported Dice scores are single runs with no standard deviations, confidence intervals, or significance tests. On the fundus dataset the target domains contain only 30 and 20 images, and on prostate many per-site differences between FA-SAM and SAMMed are small (e.g., site D 84.60 vs 82.40), while site E actually favors SAMMed. Without repeated-seed variability or a statistical test, the headline 4.99% average improvement over SAMMed cannot be distinguished from noise. Please provide at least three seeds and report mean ± std, and perform a paired significance test where the same test images are used.
  3. [Section IV-C and Table I] The text states that 'FA-SAM achieved the best overall performance in all unseen domains' and 'Our method outperformed all others,' but Table I shows that on site E, SAMMed scores 80.47% while FA-SAM scores 75.98%, and in Table II on site C, RAS4DG scores 72.88% while FA-SAM scores 71.67%. The claims should be restricted to average Dice or to the specific domains where the improvement holds. This is not a purely cosmetic issue: the paper's argument for robustness across unseen sites rests on the per-domain consistency that the data do not support.
  4. [Section III-B] The core assumption that source-to-target domain shift is representable by Gaussian/Poisson perturbations of channel-wise feature statistics is not validated, and the evidence offered in Table V is affected by the target-domain selection problem described above. Please provide a direct test of this assumption, for example by measuring the feature-statistic discrepancy between source and target encoders, or by showing that SUFM improves performance on a target domain that was not used in selecting the noise model. Absent such a validation, the mechanism behind the proposed uncertainty modeling remains unsupported.
minor comments (6)
  1. [Section III-B, Eqs. (1)–(5)] The notation is inconsistent: in Eq. (1) the feature f is written without indices, and the same symbol x is used both for the input image and for the argument of µ(x) and σ(x); Eq. (5) appears to normalize the feature map x using the original statistics but the notation is ambiguous. Please rewrite with consistent subscripts, e.g., f_bchw and normalized features, and define all variables.
  2. [Section III-D and Fig. 2] The IPEF module is repeatedly described as being inserted into the 'SAM mask encoder' (e.g., 'we introduced an Image-Prompt Embedding Fusion (IPEF) module into the SAM mask encoder' in the first paragraph and the caption of Fig. 2(B)), while the architecture and context indicate the mask decoder. Please correct the terminology throughout.
  3. [Section IV-B] The phrase 'the Adam decay optimizer' is not a standard optimizer name; please specify whether AdamW or Adam with weight decay is used, and clarify whether '200 iterations' and '100 iterations' refer to epochs or to gradient-descent iterations.
  4. [Table III] The checkmark pattern under the header (SUFM, AGM∗, IPEF, SAM) is ambiguous; the first row appears to combine SUFM with the base AGM and SAM, but it is marked only with two checkmarks. Please label each configuration explicitly.
  5. [Tables IV and V] The terms 'Preliminary Prediction' and 'Final Prediction' are used without a definition in the main text. Please define these as the AGM-only output and the final SAM output, respectively, and state what is averaged for the reported numbers.
  6. [Section III-C] The pipeline relies on taking the largest connected component of the AGM mask; please state explicitly that this post-processing is applied automatically, and discuss the case of disconnected or multiple target structures where the largest-component heuristic may fail.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FA-SAM's derivation chain is not self-referential, and its components are tested against external benchmarks; the target-domain leakage in hyperparameter selection is an evaluation-protocol concern, not a circular derivation.

full rationale

The paper's claimed derivation chain is not circular. The AGM branch with SUFM is trained on the source domain with a supervised loss and generates bounding-box prompts; the IPEF module fuses SAM image and prompt embeddings in the mask decoder; the final Dice scores are measured against external baselines on unseen target sites. No prediction is constructed by definition from a fitted quantity: SUFM perturbs source feature statistics with uncertainty sampling and Poisson noise, and the AGM is a separate trained network, not an inverse of the reported metric. The comparison to SAMMed, DeSAM, MedSAM, and other published methods provides independent external grounding, and the only author self-citations are routine context references, not load-bearing premises. The main validity concern is that Tables IV and V select the SUFM insertion position ('0-1') and the combined Gaussian+Poisson distribution using Dice on the target domains B-F, so the headline 84.53% average is partly a selection artifact and may overstate clean source-only generalization. That is a serious fairness and leakage issue, and a source-domain validation split would strengthen the evaluation, but it is not a mathematical circularity: the reported performance is not equal by construction to an input, and no fitted parameter is renamed as an independent prediction. The paper also honestly acknowledges in Table VI that FA-SAM trails SAM with ground-truth prompts, which further indicates the evaluation is not being presented as a tautological win.

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

The method introduces no physical free parameters. The listed choices are design decisions that affect the central result. The principal risks are selecting the SUFM insertion position on target-domain performance and linking noise models to domain shift without validating them.

free parameters (4)
  • SUFM insertion position = between Layer 0 and Layer 1 (Convolution Blocks 0 and 1)
    Selected as the best setting in Table IV using target-domain Dice scores on the prostate benchmark, so the design choice is tuned to the test domains rather than a held-out validation split.
  • Number of AGM and SAM training iterations = 200 and 100 iterations respectively
    Chosen by hand with no convergence analysis or validation curve; if these are iterations rather than epochs, the reported training budget is unusually small.
  • Noise model for SUFM = Gaussian plus Poisson
    Chosen over Gaussian-only and Poisson-only in Table V; selection is based on target-domain Dice, so it is a target-informed design choice rather than a derived constant.
  • Multi-scale encoder channels for IPEF = 256, 128, 64
    Hand-selected channels from SAM image encoder intermediate features; no ablation over alternative channel sets is reported.
assumptions (4)
  • domain assumption Domain shift in medical images is primarily captured by shallow feature statistics such as color, texture, and contrast, so perturbing those statistics approximates unseen target distributions.
    Invoked in Section III-B and motivated by refs [29,30], but no evidence is provided that the synthesized distributions match the actual target feature statistics.
  • ad hoc to paper The largest connected component of the AGM's preliminary mask contains the target structure, so all other components can be discarded before box-prompt generation.
    Introduced in Section III-C via BFS post-processing; plausible for single-organ prostate but unverified for multi-component vessel trees in fundus images.
  • domain assumption Poisson plus Gaussian noise is an appropriate model for real medical imaging noise in feature-statistics space.
    Section III-B justifies this with a PET reference [31], but the mapping from pixel-level noise to feature-statistics noise is not demonstrated.
  • domain assumption A Transformer-based SAM mask decoder neglects local and high-frequency details and therefore benefits from skip connections and multi-scale fusion.
    Stated in Section III-D without quantitative evidence; this motivates the IPEF module.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fully Automated SAM for Single-source Domain Generalization in Medical Image Segmentation." pith.science (2026). https://pith.science/paper/ZK37E72Q

@misc{pith2026250717281,
  author       = {Pith},
  title        = {Pith review of: Fully Automated SAM for Single-source Domain Generalization in Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZK37E72Q}},
  note         = {Machine review of arXiv:2507.17281}
}
read the original abstract

Although SAM-based single-source domain generalization models for medical image segmentation can mitigate the impact of domain shift on the model in cross-domain scenarios, these models still face two major challenges. First, the segmentation of SAM is highly dependent on domain-specific expert-annotated prompts, which prevents SAM from achieving fully automated medical image segmentation and therefore limits its application in clinical settings. Second, providing poor prompts (such as bounding boxes that are too small or too large) to the SAM prompt encoder can mislead SAM into generating incorrect mask results. Therefore, we propose the FA-SAM, a single-source domain generalization framework for medical image segmentation that achieves fully automated SAM. FA-SAM introduces two key innovations: an Auto-prompted Generation Model (AGM) branch equipped with a Shallow Feature Uncertainty Modeling (SUFM) module, and an Image-Prompt Embedding Fusion (IPEF) module integrated into the SAM mask decoder. Specifically, AGM models the uncertainty distribution of shallow features through the SUFM module to generate bounding box prompts for the target domain, enabling fully automated segmentation with SAM. The IPEF module integrates multiscale information from SAM image embeddings and prompt embeddings to capture global and local details of the target object, enabling SAM to mitigate the impact of poor prompts. Extensive experiments on publicly available prostate and fundus vessel datasets validate the effectiveness of FA-SAM and highlight its potential to address the above challenges.

Figures

Figures reproduced from arXiv: 2507.17281 by the authors.

Figure 1
Figure 1. (a) is prostate MRI collected from six different clinical centers [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (A) The overall architecture of the FA-SAM. The AGM branch with the embedded SUFM module minimizes the generation of poor bounding box prompts for the target domain, enabling fully automatic segmentation of SAM. (B) The IPEF module is introduced into the SAM mask encoder to enhance the ability of SAM to capture global context information and mitigate its sensitivity to poor prompts. B. Shallow Feature Uncertainty Mo… view at source ↗
Figure 3
Figure 3. Structure of our Shallow Feature Uncertainty Modeling ( [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of FA-SAM with state-of-the-art methods [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: This visualization result shows the AGM branch generating SAM [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Enhancing MedSAM with a Lightweight Box Predictor for Medical Image Segmentation

    cs.CV 2026-06 unverdicted novelty 5.0 of 10

    Enhances MedSAM with a 1.6M-parameter Box Predictor trained in two stages to convert single clicks to bounding boxes, reporting Dice scores of 0.89-0.98 on four medical datasets across CT, MRI, and ultrasound.

Reference graph

Works this paper leans on

45 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    Wfss: weighted fusion of spectral transformer and spatial self-attention for robust hyperspectral image classification against adversarial attacks,

    L. Tang, Z. Yin, H. Su, W. Lyu, and B. Luo, “Wfss: weighted fusion of spectral transformer and spatial self-attention for robust hyperspectral image classification against adversarial attacks,” Visual Intelligence , vol. 2, no. 1, p. 5, 2024

  2. [2]

    Correl- ative and discriminative label grouping for multi-label visual prompt tuning,

    L. Ma, S. Xu, M.-K. Xie, L. Wang, D. Sun, and H. Zhao, “Correl- ative and discriminative label grouping for multi-label visual prompt tuning,” in CVPR, 2025, pp. 25 434–25 443

  3. [3]

    Text-region matching for multi-label image recognition with missing labels,

    L. Ma, H. Xie, L. Wang, Y . Fu, D. Sun, and H. Zhao, “Text-region matching for multi-label image recognition with missing labels,” in ACM MM, 2024, pp. 6133–6142

  4. [4]

    Uav-ground visual tracking: A unified dataset and collaborative learning approach,

    D. Sun, L. Cheng, S. Chen, C. Li, Y . Xiao, and B. Luo, “Uav-ground visual tracking: A unified dataset and collaborative learning approach,” IEEE TCSVT , vol. 34, no. 5, pp. 3619–3632, 2023

  5. [5]

    Semantic-aware dual contrastive learning for multi-label image classification,

    L. Ma, D. Sun, L. Wang, H. Zhao, and B. Luo, “Semantic-aware dual contrastive learning for multi-label image classification,” in ECAI, 2023, pp. 1656–1663

  6. [6]

    Domain adaptive lung nodule detection in x-ray image,

    H. Zhao, L. Jiang, L. Ma, D. Sun, and Y . Fu, “Domain adaptive lung nodule detection in x-ray image,” in SMC. IEEE, 2024, pp. 1684– 1690

  7. [7]

    Semantic knowledge transfer for semi-supervised medical image segmentation,

    S. Zhou, H. Zhao, L. Ma, and D. Sun, “Semantic knowledge transfer for semi-supervised medical image segmentation,” EAAI, vol. 158, p. 111235, 2025

  8. [8]

    High-order paired- aspp for deep semantic segmentation networks,

    X. Sun, Y . Zhang, C. Chen, S. Xie, and J. Dong, “High-order paired- aspp for deep semantic segmentation networks,” Information Sciences, vol. 646, p. 119364, 2023

Show all 45 references
  1. [9]

    Domain generalization: A survey,

    K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy, “Domain generalization: A survey,” IEEE TPAMI , vol. 45, no. 4, pp. 4396– 4415, 2022

  2. [10]

    Rethinking data augmentation for single-source domain generalization in medical image segmentation,

    Z. Su, K. Yao, X. Yang, K. Huang, Q. Wang, and J. Sun, “Rethinking data augmentation for single-source domain generalization in medical image segmentation,” in AAAI, vol. 37, 2023

  3. [11]

    Exact feature distribu- tion matching for arbitrary style transfer and domain generalization,

    Y . Zhang, M. Li, R. Li, K. Jia, and L. Zhang, “Exact feature distribu- tion matching for arbitrary style transfer and domain generalization,” in WACV, 2022, pp. 8035–8045

  4. [12]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo et al. , “Segment anything,” in ICCV, 2023, pp. 4015–4026

  5. [13]

    A comprehensive survey on segment anything model for vision and beyond,

    C. Zhang, L. Liu, Y . Cui, G. Huang, W. Lin, Y . Yang, and Y . Hu, “A comprehensive survey on segment anything model for vision and beyond,” arXiv:2305.08196, 2023

  6. [14]

    Weaksam: Segment anything meets weakly-supervised instance-level recogni- tion,

    L. Zhu, J. Zhou, Y . Liu, X. Hao, W. Liu, and X. Wang, “Weaksam: Segment anything meets weakly-supervised instance-level recogni- tion,” in ACM MM, 2024, pp. 7947–7956

  7. [15]

    Segment anything in medical images,

    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

  8. [16]

    Sam-u: Multi-box prompts triggered uncertainty estimation for reli- able sam in medical image,

    G. Deng, K. Zou, K. Ren, M. Wang, X. Yuan, S. Ying, and H. Fu, “Sam-u: Multi-box prompts triggered uncertainty estimation for reli- able sam in medical image,” in MICCAI. Springer, 2023, pp. 368–377

  9. [17]

    Ur- sam: Enhancing the reliability of segment anything model for auto- prompting medical image segmentation with uncertainty rectification,

    Y . Zhang, S. Hu, S. Ren, T. Pan, C. Jiang, Y . Cheng, and Y . Qi, “Ur- sam: Enhancing the reliability of segment anything model for auto- prompting medical image segmentation with uncertainty rectification,” Available at SSRN 4878606 , 2023

  10. [18]

    Leveraging sam for single-source domain generalization in medical image segmentation,

    H. Wang, H. Ye, Y . Xia, and X. Zhang, “Leveraging sam for single-source domain generalization in medical image segmentation,” arXiv:2401.02076, 2024

  11. [19]

    Desam: Decoupled segment anything model for generalizable medical image segmenta- tion,

    Y . Gao, W. Xia, D. Hu, W. Wang, and X. Gao, “Desam: Decoupled segment anything model for generalizable medical image segmenta- tion,” in MICCAI. Springer, 2024, pp. 509–519

  12. [20]

    False negative/positive control for sam on noisy medical images,

    X. Yao, H. Liu, D. Hu, D. Lu, A. Lou, H. Li, R. Deng, G. Arenas, B. Oguz, N. Schwartz et al. , “False negative/positive control for sam on noisy medical images,” arXiv:2308.10382, 2023

  13. [21]

    Samus: Adapting segment anything model for clinically-friendly and general- izable ultrasound image segmentation,

    X. Lin, Y . Xiang, L. Zhang, X. Yang, Z. Yan, and L. Yu, “Samus: Adapting segment anything model for clinically-friendly and general- izable ultrasound image segmentation,” arXiv:2309.06824, 2023

  14. [22]

    Skinsam: Empowering skin can- cer segmentation with segment anything model. arxiv 2023,

    M. Hu, Y . Li, and X. Yang, “Skinsam: Empowering skin can- cer segmentation with segment anything model. arxiv 2023,” arXiv:2304.13973, 2023

  15. [23]

    Segment anything model for med- ical image segmentation: Current applications and future directions,

    Y . Zhang, Z. Shen, and R. Jiao, “Segment anything model for med- ical image segmentation: Current applications and future directions,” Computers in Biology and Medicine , p. 108238, 2024

  16. [24]

    Causality-inspired single-source domain generalization for medical image segmentation,

    C. Ouyang, C. Chen, S. Li, Z. Li, C. Qin, W. Bai, and D. Rueckert, “Causality-inspired single-source domain generalization for medical image segmentation,” IEEE TMI, vol. 42, no. 4, pp. 1095–1106, 2022

  17. [25]

    Maxstyle: Adversarial style composition for robust medical image segmentation,

    C. Chen, Z. Li, C. Ouyang, M. Sinclair, W. Bai, and D. Rueckert, “Maxstyle: Adversarial style composition for robust medical image segmentation,” in MICCAI. Springer, 2022, pp. 151–161

  18. [26]

    Robust and generalizable visual representation learning via random convolutions,

    Z. Xu, D. Liu, J. Yang, C. Raffel, and M. Niethammer, “Robust and generalizable visual representation learning via random convolutions,” arXiv:2007.13003, 2020

  19. [27]

    Domain generalization with mixstyle,

    K. Zhou, Y . Yang, Y . Qiao, and T. Xiang, “Domain generalization with mixstyle,” arXiv:2104.02008, 2021

  20. [28]

    Uncertainty modeling for out-of-distribution generalization,

    X. Li, Y . Dai, Y . Ge, J. Liu, Y . Shan, and L. Y . Duan, “Uncertainty modeling for out-of-distribution generalization,” arXiv e-prints , 2022

  21. [29]

    From shallow feature learning to deep learning: Benefits from the width and depth of deep architec- tures,

    G. Zhong, X. Ling, and L.-N. Wang, “From shallow feature learning to deep learning: Benefits from the width and depth of deep architec- tures,” Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 9, no. 1, p. e1255, 2019

  22. [30]

    Meganet: Multi-scale edge-guided attention network for weak bound- ary polyp segmentation,

    N.-T. Bui, D.-H. Hoang, Q.-T. Nguyen, M.-T. Tran, and N. Le, “Meganet: Multi-scale edge-guided attention network for weak bound- ary polyp segmentation,” in WACV, 2024, pp. 7985–7994

  23. [31]

    Positron emission tomography,

    G. Muehllehner and J. S. Karp, “Positron emission tomography,” Physics in Medicine & Biology , vol. 51, no. 13, p. R117, 2006

  24. [32]

    Asps: Aug- mented segment anything model for polyp segmentation,

    H. Li, D. Zhang, J. Yao, L. Han, Z. Li, and J. Han, “Asps: Aug- mented segment anything model for polyp segmentation,” in MICCAI. Springer, 2024, pp. 118–128

  25. [33]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

  26. [34]

    Shape-aware meta-learning for gen- eralizing prostate mri segmentation to unseen domains,

    Q. Liu, Q. Dou, and P. A. Heng, “Shape-aware meta-learning for gen- eralizing prostate mri segmentation to unseen domains,” in MICCAI. Springer, 2020

  27. [35]

    Ridge-based vessel segmentation in color images of the retina

    J. Staal, M. Abramoff, M. Niemeijer, M. A. Viergever, and B. Van Gin- neken, “Ridge-based vessel segmentation in color images of the retina.” IEEE TMI , vol. 23, no. 4, pp. 501–509, 2004

  28. [36]

    Automatic detection of vascular bifurcations and crossings in retinal images using orientation scores,

    S. Abbasi-Sureshjani, I. Smit-Ockeloen, E. Bekkers, B. Dashtbozorg, and B. ter Haar Romeny, “Automatic detection of vascular bifurcations and crossings in retinal images using orientation scores,” in ISBI, 2016, pp. 189–192

  29. [37]

    Towards a glaucoma risk index based on simulated hemodynamics from fundus images,

    J. I. Orlando, J. Barbosa Breda, K. Van Keer, M. B. Blaschko, P. J. Blanco, and C. A. Bulant, “Towards a glaucoma risk index based on simulated hemodynamics from fundus images,” in MICCAI. Springer, 2018, pp. 65–73

  30. [38]

    Self-challenging improves cross-domain generalization,

    Z. Huang, H. Wang, E. P. Xing, and D. Huang, “Self-challenging improves cross-domain generalization,” in ECCV. Springer, 2020, pp. 124–140

  31. [39]

    Realistic adversarial data augmentation for mr image segmentation,

    C. Chen, C. Qin, H. Qiu, C. Ouyang, S. Wang, L. Chen, G. Tarroni, W. Bai, and D. Rueckert, “Realistic adversarial data augmentation for mr image segmentation,” in MICCAI. Springer, 2020, pp. 667–677

  32. [40]

    Cutmix: Regularization strategy to train strong classifiers with localizable features,

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo, “Cutmix: Regularization strategy to train strong classifiers with localizable features,” in WACV, 2019, pp. 6023–6032

  33. [41]

    Deep symmetric adaptation network for cross-modality medical image seg- mentation,

    X. Han, L. Qi, Q. Yu, Z. Zhou, Y . Zheng, Y . Shi, and Y . Gao, “Deep symmetric adaptation network for cross-modality medical image seg- mentation,” IEEE TMI , vol. 41, no. 1, pp. 121–132, 2021

  34. [42]

    Generalizable cross- modality medical image segmentation via style augmentation and dual normalization,

    Z. Zhou, L. Qi, X. Yang, D. Ni, and Y . Shi, “Generalizable cross- modality medical image segmentation via style augmentation and dual normalization,” CVPR, 2021

  35. [43]

    Semantic-aware domain generalized segmentation,

    D. Peng, Y . Lei, M. Hayat, Y . Guo, and W. Li, “Semantic-aware domain generalized segmentation,” in CVPR, 2022

  36. [44]

    Frequency- mixed single-source domain generalization for medical image segmen- tation,

    H. Li, H. Li, W. Zhao, H. Fu, X. Su, Y . Hu, and J. Liu, “Frequency- mixed single-source domain generalization for medical image segmen- tation,” MICCAI, 2023

  37. [45]

    Medical image segmentation via single-source domain generalization with random amplitude spectrum synthesis,

    Q. Qiao, W. Wang, Q. Meixia, S. Kun, J. Bin, and G. Qiang, “Medical image segmentation via single-source domain generalization with random amplitude spectrum synthesis,” in MICCAI, 2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.