Pith. sign in

REVIEW 3 major objections 4 minor 36 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

Edge-aware prompts fix SAM's ultrasound boundary problem

2026-07-09 16:23 UTC pith:BNFLS43L

load-bearing objection Solid SAM adaptation for ultrasound segmentation, but the 'synergy' claim is undermined by the paper's own ablation numbers the 3 major comments →

arxiv 2607.07240 v1 pith:BNFLS43L submitted 2026-07-08 cs.CV

An Edge-aware Prompt-enhanced SAM for Ultrasound Image Segmentation

Wenhao Li , Fangyi Liu , Bo Du This is my paper
classification cs.CV
keywords ultrasound image segmentationSegment Anything Modeledge-aware segmentationprompt enhancementboundary delineationmedical image segmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The Segment Anything Model (SAM), despite its success on natural images, produces blurry or drifting boundaries when applied to ultrasound images—a modality plagued by speckle noise, low contrast, and gradual tissue transitions. This paper proposes EP-SAM, which adapts SAM by adding two modules that connect the model's image encoder to its prompt encoder. The first, the Edge-Aware Module (EAM), extracts boundary cues from intermediate features of SAM's image encoder and supervises them against ground-truth edge maps. The second, the Prompt Enhanced Module (PEM), fuses these edge-aware features with multi-level semantic features to generate coarse mask predictions, which are then fed back into SAM's prompt encoder as dense mask-form prompts. The central claim is that closing the loop between what the image encoder sees and what the prompt encoder receives—specifically by injecting explicitly supervised boundary information—resolves the boundary ambiguity that causes SAM to fail on ultrasound data. Experiments on six ultrasound benchmarks show EP-SAM outperforming prior SAM-based methods under both prompt-free and single-point-prompt settings.

Core claim

The paper identifies that SAM's failure on ultrasound images stems from treating the image encoder, prompt encoder, and mask decoder as isolated modules, causing boundary cues present in intermediate encoder features to be lost before they reach the prompt stage. By extracting edge features from intermediate transformer blocks, supervising them with ground-truth boundaries, and feeding the result back as dense mask prompts, the model achieves boundary-informed segmentation without requiring external prompt sources or heavy architectural changes.

What carries the argument

The Edge-Aware Module (EAM), which applies gated residual convolutional refinement to intermediate SAM transformer features and supervises them against Canny-extracted ground-truth edges; and the Prompt Enhanced Module (PEM), which fuses multi-level encoder features with EAM outputs through dynamic gating to produce coarse masks that serve as dense prompts for SAM's frozen prompt encoder.

Load-bearing premise

The paper assumes that the boundary features extracted from intermediate transformer blocks provide information that is complementary to, rather than redundant with, what the mask decoder already processes, and that the observed gains stem from the coupling mechanism rather than simply from adding more supervised training signals.

What would settle it

If feeding the same edge-aware features directly to the mask decoder (bypassing the prompt encoder loop) yields equivalent or better results, the claim that prompt-level integration is the key mechanism would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If intermediate encoder features in SAM carry boundary information that the mask decoder cannot recover on its own, then any SAM adaptation for modalities with ambiguous boundaries (not just ultrasound) could benefit from explicit edge supervision at the encoder level.
  • Self-generated dense mask prompts derived from internal features may substitute for external prompting (points, boxes, or language models), reducing the need for user interaction or large language model integration in medical imaging workflows.
  • The principle of bridging encoder and prompt encoder with task-specific intermediate supervision could generalize to other foundation-model adaptation scenarios where the pre-trained model's components are under-coupled.

Where Pith is reading between the lines

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

  • If the performance gain comes primarily from the additional supervision signal rather than from the architectural coupling between encoder and prompt encoder, then simply adding an auxiliary edge-loss branch to the mask decoder might achieve similar gains without the prompt-generation machinery.
  • The claim that EAM and PEM improve 'synergy' between SAM components would be strengthened by showing that removing the feedback loop (feeding edge features directly to the decoder rather than through the prompt encoder) degrades performance—if it does not, the synergy claim is weaker than the architectural narrative suggests.
  • The use of Canny-extracted edges as ground truth introduces a dependency on classical edge detection; if the quality of these pseudo-labels varies across anatomical structures, the method's boundary improvement may be uneven across datasets in ways not captured by aggregate Dice scores.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The manuscript proposes EP-SAM, an adaptation of the Segment Anything Model (SAM) for ultrasound image segmentation. The method introduces two modules: an Edge-Aware Module (EAM) that extracts boundary cues from intermediate features of SAM's image encoder under explicit edge supervision, and a Prompt Enhanced Module (PEM) that fuses these edge-aware features with multi-block semantic features to generate dense mask-form prompts for the prompt encoder. The approach is evaluated on six ultrasound datasets (three in-domain, three cross-domain) against 14 baselines under prompt-free and single-point-prompt settings, with ablation studies isolating the contributions of EAM and PEM.

Significance. Ultrasound image segmentation suffers from boundary ambiguity and speckle noise, and adapting SAM to this domain is an active area of research. The paper's approach of using intermediate encoder features to generate self-prompts, rather than relying on external models or sparse user-provided prompts, is a reasonable architectural contribution. The experimental evaluation is relatively thorough, covering multiple anatomical structures and including both in-domain and cross-domain generalization tests. The inclusion of ablation studies (Tables II and III) to isolate component contributions is a positive practice.

major comments (3)
  1. §IV.C, Table II: The paper's central conceptual claim is that EAM and PEM are 'complementary' and produce 'enhanced synergy' between SAM's image and prompt encoders. However, the ablation results show sub-additive gains when the two modules are combined. On TN3K, EAM alone gives +1.22 Dice and PEM alone gives +1.92, but combined gives +2.47 (vs. 3.14 if additive). On BUSI, EAM gives +1.14 and PEM gives +3.23, but combined gives +3.78 (vs. 4.37 if additive). This pattern indicates overlapping or redundant information capture rather than complementarity. The paper should either temper the 'synergy/complementary' framing to match the evidence or provide additional analysis (e.g., feature similarity analysis, gradient-weighted class activation maps) demonstrating that the two modules capture genuinely non-redundant information. As stated, the mechanistic claim is not supported by the paper's
  2. §IV.B, Table I: The absolute improvements over the strongest baselines are small (e.g., +0.61 Dice over UN-SAM without prompts on average, +0.13 Dice over CC-SAM with point prompts on average). No statistical significance tests (e.g., p-values, confidence intervals, or multiple random seed runs) are reported. Given the modest margins, it is unclear whether these differences are robust or within noise. The authors should add significance testing or report variance across multiple runs to substantiate the claim of consistent outperformance.
  3. §III.A, §III.C: The EAM introduces additional supervision via the edge loss (L_edge), and PEM introduces the coarse mask loss (L_coarse). The ablation in Table II does not isolate whether the performance gains come from the architectural modules themselves or simply from the addition of auxiliary supervised signals. A fairer test of the 'synergy' claim would include a baseline that adds equivalent auxiliary supervision signals (e.g., a simple edge prediction head and a coarse mask prediction head) without the proposed cross-module feature routing, to verify that the gains are attributable to the architectural design rather than the multi-task supervision alone.
minor comments (4)
  1. §IV.A.1: The dataset split ratios for BUSI are rendered as unreadable Unicode escape sequences (e.g., '/uni00000038/uni00000010/...'). This should be corrected to the standard '7:1:2' format.
  2. §III.C: The loss term weights for L_edge, L_coarse, and L_final are all set to 1, and α is fixed to 0.8. No sensitivity analysis is provided for these hyperparameters. A brief discussion of robustness to these choices would strengthen the paper.
  3. Figure 2: The diagram is dense and some data flow paths (e.g., how the dynamic gating weights w1–w4 are computed and applied) are difficult to follow. Clarifying the flow or splitting the figure would aid readability.
  4. §IV.B: The comparison with CC-SAM (an LLM-assisted method) is mentioned in the text but CC-SAM is only evaluated on the average column in Table I, not on individual datasets. For completeness, per-dataset results for CC-SAM should be included or the comparison scope should be clarified.

Circularity Check

0 steps flagged

No circularity: EP-SAM is a standard deep learning method evaluated against external benchmarks with standard metrics

full rationale

The paper proposes EP-SAM, an edge-aware and prompt-enhanced adaptation of SAM for ultrasound image segmentation. The method consists of two modules: an Edge-Aware Module (EAM) that extracts boundary cues from intermediate transformer features and supervises them with ground-truth-derived edge maps (via Canny), and a Prompt Enhanced Module (PEM) that fuses multi-block features to generate coarse mask prompts fed into SAM's prompt encoder. The derivation chain is entirely self-contained: (1) edge ground truth is extracted from segmentation masks via Canny edge detection — a standard preprocessing step, not a circular definition; (2) the self-prompting mechanism generates prompts from the model's own intermediate features, but these are supervised by ground-truth labels during training (L_coarse and L_final in Eq. 4), so the prompt generation is not defined in terms of the prediction output; (3) the loss function (Eq. 3-4) combines Dice and BCE losses applied to edge maps, coarse masks, and final segmentation — all supervised against ground truth, none defined in terms of each other; (4) evaluation uses standard external benchmarks (BUSI, TN3K, CAMUS, DDTI, UDIAT, HMC-QU) with Dice and HD metrics. The skeptic's concern about sub-additive gains in Table II is a correctness/interpretation issue (whether 'synergy' is the right framing), not a circularity issue — the ablation numbers are not constructed to equal their inputs by definition. No self-citation chain is load-bearing for the central claims. The method is evaluated against external benchmarks with standard metrics, and no prediction reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

The axiom ledger captures the key assumptions and designed components. The free parameters are standard hyperparameters without deep justification but are consistent with prior work. The core axiom — that intermediate SAM features contain extractable boundary information — is plausible and partially validated by the ablation results, though the causal mechanism is not isolated.

free parameters (3)
  • α (loss weight for BCE vs Dice) = 0.8
    Fixed hyperparameter controlling the balance between BCE and Dice loss in Eq. (3). Stated as following SAMUS [19] without independent justification.
  • Loss term weights (L_edge, L_coarse, L_final) = 1, 1, 1
    All three supervision terms in Eq. (4) are set to equal weight 1 without sensitivity analysis or justification.
  • Number of selected transformer blocks (n) = 4
    Four Global Transformer Block features (g1–g4) are used. Table III ablates this choice but the default configuration uses all four.
axioms (3)
  • domain assumption SAM's intermediate transformer features contain boundary-relevant information that can be extracted by lightweight convolutional blocks.
    §III.A introduces EAM on the premise that intermediate features g_i preserve fine-grained spatial structures. This is plausible but not independently verified before use.
  • domain assumption Canny edge detection on ground-truth masks provides a valid boundary supervision signal.
    §III.C states the ground-truth boundary map for L_edge is extracted via Canny edge detection. This is a standard but heuristic choice that affects what the model learns.
  • domain assumption Mona adapters [29] provide superior feature modulation compared to linear adapters for ultrasound.
    §IV.C briefly mentions Mona adapters achieve 81.61% vs 80.23% Dice on BUSI, but this is stated in passing without full experimental detail.
invented entities (2)
  • Edge-Aware Block (EAB) independent evidence
    purpose: Residual convolutional refinement + gated spatial interaction to extract boundary features from intermediate SAM encoder features.
    The EAB is a new architectural component. Its effectiveness is tested via ablation (Table II) and qualitative results (Figure 4). It is not a postulated physical entity but a designed module with falsifiable performance.
  • Prompt Enhanced Module (PEM) independent evidence
    purpose: Fuses multi-level encoder features with edge-aware outputs to generate coarse mask prompts for SAM's prompt encoder.
    PEM is a new module. Its contribution is isolated in ablation (Table II) and its multi-block design is ablated in Table III. Falsifiable through performance metrics.

pith-pipeline@v1.1.0-glm · 13683 in / 2469 out tokens · 367119 ms · 2026-07-09T16:23:48.034435+00:00 · methodology

0 comments
read the original abstract

Ultrasound image segmentation is essential for delineating anatomical structures and lesions, providing the foundation for accurate diagnosis. While the Segment Anything Model (SAM) has demonstrated remarkable success on natural images, its performance on ultrasound data is often hindered by poor boundary delineation. To address this limitation, we propose EP-SAM, an edge-aware and prompt-enhanced adaptation of SAM. Specifically, we leverage multi-block feature extraction from the image encoder to enrich coarse-to-fine semantic representations, while edge-aware supervision of the image encoder improves robustness to contour ambiguity and speckle noise. By integrating these complementary cues, EP-SAM generates high-quality prompts that effectively guide the model toward target regions of interest. Experimental results on multiple benchmarks demonstrate that EP-SAM consistently outperforms existing SAM-based methods.

Figures

Figures reproduced from arXiv: 2607.07240 by Bo Du, Fangyi Liu, Wenhao Li.

Figure 1
Figure 1. Figure 1: Comparison of SAM-based adaptations for medical image [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of EP-SAM with two key modules: an edge-aware module that enhances contour modeling and a prompt enhanced module [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Generalization performance (Dice %) on three unseen ultra [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of EP-SAM and other SOTA methods. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

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

36 extracted references · 36 canonical work pages · 4 internal anchors

  1. [1]

    Deep learning approaches for data augmentation and classification of breast masses using ultrasound images,

    W. Al-Dhabyaniet al., “Deep learning approaches for data augmentation and classification of breast masses using ultrasound images,”Int. J. Adv. Comput. Sci. Appl, 2019

    work page 2019

  2. [2]

    Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network,

    M. Byraet al., “Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network,”BSPC, 2020

    work page 2020

  3. [3]

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

    H. Caoet al., “Swin-unet: Unet-like pure transformer for medical image segmentation,” inECCV, 2022

    work page 2022

  4. [4]

    Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images,

    G. Chenet al., “Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images,”TMI, 2022

    work page 2022

  5. [5]

    Transunet: Rethinking the u-net architecture de- sign for medical image segmentation through the lens of transformers,

    J. Chen, J. Meiet al., “Transunet: Rethinking the u-net architecture de- sign for medical image segmentation through the lens of transformers,” MIA, 2024

    work page 2024

  6. [6]

    Ba-sam: Boundary-aware adaptation of segment any- thing model for medical image segmentation,

    Y . Chenet al., “Ba-sam: Boundary-aware adaptation of segment any- thing model for medical image segmentation,” inBIBM, 2024

    work page 2024

  7. [7]

    Un-sam: Domain-adaptive self-prompt segmentation for universal nuclei images,

    Z. Chenet al., “Un-sam: Domain-adaptive self-prompt segmentation for universal nuclei images,”MIA, 2025

    work page 2025

  8. [8]

    Early myocardial infarction detection over multi-view echocardiography,

    A. Degerliet al., “Early myocardial infarction detection over multi-view echocardiography,”BSPC, 2024

    work page 2024

  9. [9]

    Cpfnet: Context pyramid fusion network for medical image segmentation,

    S. Fenget al., “Cpfnet: Context pyramid fusion network for medical image segmentation,”TMI, 2020

    work page 2020

  10. [10]

    Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules,

    H. Gonget al., “Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules,”CBM, 2023

    work page 2023

  11. [11]

    Cc-sam: Sam with cross-feature attention and context for ultrasound image segmentation,

    S. N. Gowdaet al., “Cc-sam: Sam with cross-feature attention and context for ultrasound image segmentation,” inECCV, 2024

    work page 2024

  12. [12]

    Ca-net: Comprehensive attention convolutional neural networks for explainable medical image segmentation,

    R. Guet al., “Ca-net: Comprehensive attention convolutional neural networks for explainable medical image segmentation,”TMI, 2020

    work page 2020

  13. [13]

    Ce-net: Context encoder network for 2d medical image segmentation,

    Z. Guet al., “Ce-net: Context encoder network for 2d medical image segmentation,”TMI, 2019

    work page 2019

  14. [14]

    H2former: An efficient hierarchical hybrid transformer for medical image segmentation,

    A. Heet al., “H2former: An efficient hierarchical hybrid transformer for medical image segmentation,”TMI, 2023

    work page 2023

  15. [15]

    Missformer: An effective transformer for 2d medical image segmentation,

    X. Huang, Z. Denget al., “Missformer: An effective transformer for 2d medical image segmentation,”TMI, 2022

    work page 2022

  16. [16]

    Segment anything,

    A. Kirillovet al., “Segment anything,” inICCV, 2023

    work page 2023

  17. [17]

    Deep learning for segmentation using an open large- scale dataset in 2d echocardiography,

    S. Leclercet al., “Deep learning for segmentation using an open large- scale dataset in 2d echocardiography,”TMI, 2019

    work page 2019

  18. [18]

    Uniultra: Interactive parameter-efficient sam2 for universal ultrasound segmentation,

    Y . Liet al., “Uniultra: Interactive parameter-efficient sam2 for universal ultrasound segmentation,”arXiv:2511.15771, 2025

    work page arXiv 2025

  19. [19]

    Beyond adapting sam: Towards end-to-end ultrasound image segmentation via auto prompting,

    X. Linet al., “Beyond adapting sam: Towards end-to-end ultrasound image segmentation via auto prompting,” inMICCAI, 2024

    work page 2024

  20. [20]

    A review of deep-learning-based medical image segmen- tation methods,

    X. Liuet al., “A review of deep-learning-based medical image segmen- tation methods,”Sustainability, vol. 13, no. 3, p. 1224, 2021

    work page 2021

  21. [21]

    Segment anything in medical images,

    J. Maet al., “Segment anything in medical images,”NC, 2024

    work page 2024

  22. [22]

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

    O. Oktayet al., “Attention u-net: Learning where to look for the pancreas,”arXiv:1804.03999, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  23. [23]

    An open access thyroid ultrasound image database,

    L. Pedrazaet al., “An open access thyroid ultrasound image database,” in10th SIPAIM, 2015

    work page 2015

  24. [24]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronnebergeret al., “U-net: Convolutional networks for biomedical image segmentation,” inMICCAI, 2015

    work page 2015

  25. [25]

    SAM-EG: Segment Anything Model with Egde Guidance framework for efficient Polyp Segmentation

    Q.-H. Trinhet al., “Sam-eg: segment anything model with egde guidance framework for efficient polyp segmentation,”arXiv:2406.14819, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Fat-net: Feature adaptive transformers for automated skin lesion segmentation,

    H. Wuet al., “Fat-net: Feature adaptive transformers for automated skin lesion segmentation,”MIA, 2022

    work page 2022

  27. [27]

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

    J. Wuet al., “Medical sam adapter: Adapting segment anything model for medical image segmentation,”MIA, 2025

    work page 2025

  28. [28]

    Pgp-sam: Prototype-guided prompt learning for efficient few-shot medical image segmentation,

    Z. Yanet al., “Pgp-sam: Prototype-guided prompt learning for efficient few-shot medical image segmentation,” inISBI, 2025

    work page 2025

  29. [29]

    5%> 100%: Breaking performance shackles of full fine- tuning on visual recognition tasks,

    D. Yinet al., “5%> 100%: Breaking performance shackles of full fine- tuning on visual recognition tasks,” inCVPR, 2025

    work page 2025

  30. [30]

    Easam: an edge-aware sam-based paradigm for tooth segmentation,

    J. Zhanget al., “Easam: an edge-aware sam-based paradigm for tooth segmentation,”Signal, Image and Video Processing, 2025

    work page 2025

  31. [31]

    Customized Segment Anything Model for Medical Image Segmentation

    K. Zhang and D. Liu, “Customized segment anything model for medical image segmentation,”arXiv:2304.13785, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  32. [32]

    Adapting Vision Foundation Models for Real-time Ultrasound Image Segmentation

    X. Zhanget al., “Adapting vision foundation models for real-time ultrasound image segmentation,”arXiv:2503.24368, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    Segment anything model for medical image segmen- tation: Current applications and future directions,

    Y . Zhanget al., “Segment anything model for medical image segmen- tation: Current applications and future directions,”CBM, 2024

    work page 2024

  34. [34]

    Transfuse: Fusing transformers and cnns for medical image segmentation,

    Y . Zhang, H. Liu, and Q. Hu, “Transfuse: Fusing transformers and cnns for medical image segmentation,” inMICCAI, 2021

    work page 2021

  35. [35]

    Rethinking semantic segmentation from a sequence- to-sequence perspective with transformers,

    S. Zhenget al., “Rethinking semantic segmentation from a sequence- to-sequence perspective with transformers,” inCVPR, 2021

    work page 2021

  36. [36]

    Unet++: A nested u-net architecture for medical image segmentation,

    Z. Zhouet al., “Unet++: A nested u-net architecture for medical image segmentation,” inDLMIA and ML-CDS 2018, 2018

    work page 2018