Pith. sign in

REVIEW 4 major objections 5 minor 46 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 · grok-4.5

OBBSeg turns oriented bounding boxes into near full-supervision accuracy for irregular lesion segmentation across 13 medical datasets.

2026-07-08 20:04 UTC pith:3SVQD3W2

load-bearing objection Practical OBB weak-supervision stack for irregular lesions; the near-full-supervision claim likely rides more on PAFE/DBFE than on OBB geometry alone. the 4 major comments →

arxiv 2607.06007 v1 pith:3SVQD3W2 submitted 2026-07-07 cs.CV

OBBSeg: Irregular Lesion Segmentation under Oriented Bounding Box Annotations

Jun Wei , Xinchang Liu , Yu Liu , Chuhua Yang , Shuhui Wang , Hui Huang This is my paper
classification cs.CV
keywords oriented bounding boxesweak supervisionmedical image segmentationlesion segmentationgeometric consistencyMask-to-OBB lossprompt-driven modulesirregular lesions
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.

Pixel-level labels remain expensive in medical image segmentation, so weak supervision is attractive but usually too loose for irregular lesions. This paper introduces OBBSeg, an intermediate paradigm that uses oriented bounding boxes to supply both spatial extent and orientation, better matching elongated or anisotropic lesions than ordinary axis-aligned boxes. A differentiable Mask-to-OBB loss forces predicted masks to stay geometrically consistent with those boxes and thereby reduces rectangular bias, while two prompt modules (PAFE and DBFE) sharpen foreground cues and suppress background. Experiments on 13 datasets spanning five imaging modalities show that the approach outperforms prior weakly supervised methods and reaches accuracy comparable to fully supervised training. The result offers a practical route to scale medical segmentation without dense pixel painting.

Core claim

Oriented bounding boxes, when paired with a differentiable Mask-to-OBB geometric consistency loss and prompt-driven semantic modules, supply enough shape and orientation constraint to recover irregular lesion boundaries at a level comparable to full pixel-level supervision while remaining far cheaper to annotate.

What carries the argument

The Mask-to-OBB loss: a differentiable geometric consistency term that maps the predicted mask back onto the oriented-box region so the network cannot collapse to a rectangular bias. Two complementary prompt modules (PAFE for foreground enhancement and DBFE for background suppression) add semantic guidance that further tightens the mask.

Load-bearing premise

That oriented boxes plus the geometric consistency loss already give enough shape constraint to recover irregular lesion boundaries without leftover rectangular bias or heavy dependence on the semantic prompt modules.

What would settle it

On a held-out collection of highly concave or multi-component lesions, measure whether OBBSeg Dice remains well below a fully supervised baseline while residual rectangular artifacts persist inside the predicted masks; a large, systematic gap would falsify the claim that OBB geometry is a tight enough intermediate.

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

If this is right

  • Clinical pipelines can replace dense pixel painting with cheaper OBB drawing for many irregular-lesion tasks while retaining near-full-supervision accuracy.
  • Weakly supervised methods that currently rely only on axis-aligned boxes can be upgraded by switching to oriented boxes plus geometric consistency.
  • The same recipe generalizes across at least five imaging modalities and thirteen datasets without modality-specific redesign.
  • Annotation budgets can be redirected from exhaustive masks toward rapid OBB labels, lowering the cost of building large medical segmentation datasets.

Where Pith is reading between the lines

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

  • If a single OBB is already intermediate enough, similar oriented-box supervision may transfer to non-medical elongated structures such as vessels, cracks or roads.
  • The residual gap to full supervision, if any, is likely largest for highly concave or multi-part lesions where one OBB under-constrains topology.
  • An active-learning loop that escalates only ambiguous OBBs to full masks could cut annotation cost still further while preserving accuracy.
  • Differentiable geometric consistency of this form could be applied to other intermediate annotations such as ellipses or sparse polygons.

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

4 major / 5 minor

Summary. The manuscript proposes OBBSeg, a weakly supervised segmentation framework that uses oriented bounding boxes (OBBs) as an intermediate supervision signal for irregular lesion segmentation. It introduces a differentiable Mask-to-OBB geometric consistency loss intended to reduce the rectangular bias of OBB labels, together with two prompt-driven modules (PAFE and DBFE) that enhance foreground representation and suppress background. The central empirical claim is that, across 13 datasets spanning five imaging modalities, OBBSeg outperforms existing weakly supervised methods and reaches performance comparable to fully supervised baselines, while requiring only OBB annotations rather than pixel masks.

Significance. If the near-full-supervision parity claim holds under fair controls, the work would be a practically useful contribution: OBB labels are cheaper than masks yet more informative than axis-aligned boxes for elongated or anisotropic lesions, and a reproducible intermediate-supervision recipe would matter for scalable medical segmentation. The paper also ships public code, which strengthens reproducibility. The significance, however, rests on whether OBB geometry plus Mask-to-OBB is genuinely the tight intermediate that recovers irregular boundaries, or whether reported parity is driven mainly by the semantic prompt modules and backbone capacity.

major comments (4)
  1. The load-bearing premise is that OBB labels plus the Mask-to-OBB loss supply enough geometric constraint to recover irregular lesion boundaries at near fully-supervised accuracy. Mask-to-OBB matches the min-OBB of the predicted mask to the annotated OBB and therefore constrains support, scale, and orientation, but is satisfied by any mask (rectangle, ellipse, or irregular) whose min-OBB equals the annotation. Irregular boundary shape is under-determined by this loss alone. The manuscript needs a clear ablation isolating (i) OBB + Mask-to-OBB without PAFE/DBFE, (ii) PAFE/DBFE without Mask-to-OBB, and (iii) the full model, reported against the fully supervised upper bound on the same splits. Without that isolation, the claim that OBB geometry is the key intermediate remains unproven and the near-full-supervision parity may be attributable to semantic prompt cues rather than OBB supervision
  2. Relatedly, residual rectangular bias is not adequately checked. Because Mask-to-OBB is satisfied by masks that fill the OBB interior, predictions may systematically over-cover relative to true irregular ground-truth masks. The paper should report quantitative bias diagnostics (e.g., predicted-mask area / OBB area vs. GT-mask area / OBB area; boundary irregularity metrics such as perimeter-to-area or fractal-like roughness; and qualitative failure cases on highly concave lesions) so that readers can judge whether the method truly recovers irregular shapes or merely produces OBB-filling blobs that score well on overlap metrics.
  3. The experimental claim of 'performance comparable to fully supervised approaches' on 13 datasets is central and must be stated with precise numbers, variance, and protocol. Report mean±std (or CI) over multiple seeds, the exact fully supervised backbone and training recipe used as the upper bound, and whether OBB labels were derived automatically from GT masks or drawn independently by annotators. Automatic min-OBB-from-GT is a common but optimistic protocol that understates real annotation noise; if that is what was used, it should be disclosed and, ideally, complemented by a human-drawn OBB subset. Statistical tests or at least paired per-dataset deltas versus the strongest weak baseline and the full-supervision band are needed before the parity claim can be accepted.
  4. The multi-loss design (Mask-to-OBB weight and related coefficients) and the PAFE/DBFE modules introduce free parameters and architectural capacity that are not free under pure OBB supervision. Sensitivity of the main results to the Mask-to-OBB loss weight, and a controlled comparison against a strong box-supervised baseline that uses the same backbone and comparable prompt capacity without OBB-specific geometry, are required to show that gains come from the OBB intermediate rather than from extra modules or tuning.
minor comments (5)
  1. Define PAFE and DBFE on first use in the abstract and introduction (Prompt-Aware / Dual-Branch Feature Enhancement or whatever the expansions are) so readers do not have to reverse-engineer acronyms.
  2. Clarify the exact differentiable construction of Mask-to-OBB (how min-area oriented rectangle is obtained through the network, any relaxations or approximations, and gradient flow) with an equation and a short derivation or pseudocode; the abstract asserts differentiability but the geometric details matter for reproducibility.
  3. List the 13 datasets and five modalities explicitly in a table with image counts, lesion types, and train/val/test splits so the multi-modality claim is auditable at a glance.
  4. When claiming outperformance over 'existing weakly supervised methods,' name the primary competitors (e.g., CAM-based, scribble, axis-aligned box methods) and ensure they are re-implemented or re-evaluated under the same backbone and splits where possible.
  5. Code is linked; please pin the commit/release used for the reported numbers and include training configs for the main tables to make the GitHub artifact match the paper.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The four major comments correctly identify where the manuscript’s central claim—that OBB geometry plus Mask-to-OBB is a tight intermediate for irregular lesions—must be isolated more rigorously from prompt capacity, residual rectangular bias, protocol details, and free parameters. We agree with the substance of each point and will revise the experimental section accordingly: new ablations that factor OBB/Mask-to-OBB from PAFE/DBFE, quantitative bias diagnostics, multi-seed statistics with explicit full-supervision protocol and OBB provenance, loss-weight sensitivity, and a same-backbone box-supervised control. We believe these additions will make the contribution clearer without changing the core method.

read point-by-point responses

  1. Referee: The load-bearing premise is that OBB labels plus the Mask-to-OBB loss supply enough geometric constraint to recover irregular lesion boundaries at near fully-supervised accuracy. Mask-to-OBB matches the min-OBB of the predicted mask to the annotated OBB and therefore constrains support, scale, and orientation, but is satisfied by any mask (rectangle, ellipse, or irregular) whose min-OBB equals the annotation. Irregular boundary shape is under-determined by this loss alone. The manuscript needs a clear ablation isolating (i) OBB + Mask-to-OBB without PAFE/DBFE, (ii) PAFE/DBFE without Mask-to-OBB, and (iii) the full model, reported against the fully supervised upper bound on the same splits. Without that isolation, the claim that OBB geometry is the key intermediate remains unproven and the near-full-supervision parity may be attributable to semantic prompt cues rather than OBB supervision

    Authors: We agree that Mask-to-OBB alone under-determines boundary irregularity: any mask whose minimum OBB matches the annotation satisfies the geometric term, so shape recovery cannot be attributed to OBB geometry without isolating the prompt modules. The current manuscript does not present the three-way factorization the referee requests, and that omission weakens the load-bearing claim. In revision we will add a controlled ablation on the same splits and backbone, reporting (i) OBB + Mask-to-OBB without PAFE/DBFE, (ii) PAFE/DBFE without Mask-to-OBB (using only standard weak box-style or prompt cues as applicable), and (iii) the full OBBSeg model, each against the identical fully supervised upper bound. We will discuss which component closes how much of the gap to full supervision and qualify the “OBB as intermediate” claim accordingly. If (i) alone remains far from the full-supervision band, we will state that OBB geometry is necessary but not sufficient and that semantic prompts carry substantial weight. revision: yes

  2. Referee: Relatedly, residual rectangular bias is not adequately checked. Because Mask-to-OBB is satisfied by masks that fill the OBB interior, predictions may systematically over-cover relative to true irregular ground-truth masks. The paper should report quantitative bias diagnostics (e.g., predicted-mask area / OBB area vs. GT-mask area / OBB area; boundary irregularity metrics such as perimeter-to-area or fractal-like roughness; and qualitative failure cases on highly concave lesions) so that readers can judge whether the method truly recovers irregular shapes or merely produces OBB-filling blobs that score well on overlap metrics.

    Authors: This concern is well founded. Overlap metrics (Dice/IoU) can remain high for OBB-filling blobs even when boundaries are systematically smoother or more rectangular than the ground truth. The manuscript currently lacks quantitative checks for residual rectangular bias. In revision we will report, on representative datasets spanning elongated and irregular lesions: (1) predicted-mask area / OBB area versus GT-mask area / OBB area (mean and distribution); (2) boundary irregularity measures such as perimeter-to-area ratio and a simple roughness/compactness index for predictions versus GT; and (3) qualitative failure cases focused on highly concave or multi-lobed lesions, including side-by-side comparison with the annotated OBB and the fully supervised prediction. If systematic over-coverage or under-irregularity remains, we will document it and temper claims that Mask-to-OBB fully removes rectangular bias. revision: yes

  3. Referee: The experimental claim of 'performance comparable to fully supervised approaches' on 13 datasets is central and must be stated with precise numbers, variance, and protocol. Report mean±std (or CI) over multiple seeds, the exact fully supervised backbone and training recipe used as the upper bound, and whether OBB labels were derived automatically from GT masks or drawn independently by annotators. Automatic min-OBB-from-GT is a common but optimistic protocol that understates real annotation noise; if that is what was used, it should be disclosed and, ideally, complemented by a human-drawn OBB subset. Statistical tests or at least paired per-dataset deltas versus the strongest weak baseline and the full-supervision band are needed before the parity claim can be accepted.

    Authors: We accept that the parity claim must be stated with stricter protocol transparency and statistics. In the revised manuscript we will: (1) report mean±std over multiple random seeds for OBBSeg, the strongest weak baselines, and the fully supervised upper bound on the same splits; (2) specify the exact fully supervised backbone, losses, and training recipe used as the upper bound so that capacity is matched; (3) disclose OBB provenance explicitly—if OBBs were obtained as minimum OBBs from GT masks (the common automatic protocol), we will state that and note that it is optimistic relative to independent human OBB annotation; where feasible we will add a small human-drawn OBB subset or a controlled noise study to bound sensitivity to annotation noise; and (4) provide paired per-dataset deltas (and, where appropriate, simple significance tests) versus the strongest weak baseline and versus the full-supervision band. We will rephrase “comparable to fully supervised” to the precise numerical relationship that the multi-seed results support, rather than a blanket parity claim. revision: yes

  4. Referee: The multi-loss design (Mask-to-OBB weight and related coefficients) and the PAFE/DBFE modules introduce free parameters and architectural capacity that are not free under pure OBB supervision. Sensitivity of the main results to the Mask-to-OBB loss weight, and a controlled comparison against a strong box-supervised baseline that uses the same backbone and comparable prompt capacity without OBB-specific geometry, are required to show that gains come from the OBB intermediate rather than from extra modules or tuning.

    Authors: We agree that free parameters and extra module capacity must be controlled before attributing gains to the OBB intermediate. In revision we will include: (1) a sensitivity study of the main metrics to the Mask-to-OBB loss weight (and related multi-loss coefficients) over a reasonable range, reporting stability or degradation; and (2) a controlled comparison against a strong axis-aligned (or generic box) supervised baseline that uses the same backbone and comparable prompt-style capacity (PAFE/DBFE or an equivalent prompt pathway) but without OBB-specific geometry or the Mask-to-OBB term. This isolates whether orientation-aware OBB supervision and Mask-to-OBB add value beyond backbone capacity and semantic prompts. Results will be discussed honestly: if the OBB-specific path yields only marginal gains under matched capacity, we will revise the narrative to emphasize the joint system rather than OBB geometry alone. revision: yes

Circularity Check

0 steps flagged

No circularity: standard empirical weakly-supervised segmentation method paper with no derivation that reduces claimed results to inputs by construction.

full rationale

OBBSeg is a conventional deep-learning methods paper: it proposes an intermediate OBB supervision paradigm, a differentiable Mask-to-OBB geometric consistency loss, and two prompt modules (PAFE/DBFE), then trains and evaluates on 13 datasets across 5 modalities against weak- and full-supervision baselines. The strongest claim is purely empirical performance (outperforms existing WSS methods; comparable to full supervision). There is no first-principles derivation, uniqueness theorem, fitted-parameter-as-prediction step, or self-definitional identity that would make any reported metric equal its inputs by construction. Self-citations (if any in the full text) are ordinary prior-work references and are not load-bearing for the central claim. Mask-to-OBB is a training loss that constrains predicted-mask min-OBB to the annotation; it does not tautologically force the Dice/IoU numbers reported at test time. Residual concerns about whether OBB geometry alone recovers irregular boundaries (vs. semantic prompts) are correctness/assumption risks, not circularity. Per the analyzer rules this is the expected honest non-finding: score 0, empty steps.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 3 invented entities

Abstract-only audit. The claim rests on standard deep-learning practice plus the design premise that OBB geometry plus a differentiable Mask-to-OBB consistency term is adequate intermediate supervision for irregular lesions. Free parameters (loss weights, architecture hyperparameters) are not disclosed in the abstract. PAFE and DBFE are proposed modules (method components), listed as invented entities only in the architectural sense; they lack independent external evidence beyond the paper’s own experiments. No physical constants or formal axioms appear.

free parameters (2)
  • Mask-to-OBB loss weight and related multi-loss coefficients
    Any weighted combination of segmentation loss and Mask-to-OBB consistency will involve coefficients that are typically tuned on validation data; values not given in the abstract.
  • PAFE/DBFE and backbone training hyperparameters
    Learning rates, schedules, prompt design, and module capacities are free design choices that affect reported accuracy; unspecified in the abstract.
axioms (3)
  • domain assumption Oriented bounding boxes provide sufficiently informative geometric supervision for irregular/anisotropic lesions relative to axis-aligned boxes.
    Core premise of the OBBSeg intermediate-supervision paradigm stated in the abstract; not independently proven, only motivated.
  • ad hoc to paper A differentiable Mask-to-OBB geometric consistency loss can mitigate rectangular bias of OBB labels enough to approach full-mask performance.
    Central methodological bet of the paper; success is empirical and not guaranteed by the abstract’s description alone.
  • domain assumption Standard supervised deep segmentation training and evaluation protocols apply (held-out test sets, Dice/IoU-style metrics, etc.).
    Implicit background for all reported comparisons to weak and full supervision.
invented entities (3)
  • Mask-to-OBB loss no independent evidence
    purpose: Differentiable geometric consistency between predicted masks and OBB regions to reduce rectangular bias.
    Named contribution of the paper; independent evidence would be external replications or theoretical guarantees, neither available from the abstract.
  • PAFE module no independent evidence
    purpose: Prompt-driven enhancement of foreground lesion representation.
    Architectural component introduced for semantic guidance; evidence is internal to the paper’s experiments.
  • DBFE module no independent evidence
    purpose: Prompt-driven suppression of background interference.
    Architectural component introduced for semantic guidance; evidence is internal to the paper’s experiments.

pith-pipeline@v0.9.1-grok · 6293 in / 2926 out tokens · 56511 ms · 2026-07-08T20:04:47.230480+00:00 · methodology

0 comments
read the original abstract

Pixel-level annotation remains a major bottleneck in medical image segmentation, making weak supervision an attractive yet under-constrained alternative. We propose OBBSeg, an intermediate supervision paradigm guided by Oriented Bounding Boxes (OBBs) that bridges the gap between full and weak supervision. By jointly encoding spatial extent and orientation, OBBs provide compact geometric supervision that better aligns with elongated or anisotropic lesions, reducing the ambiguity of coarse box annotations. To mitigate the inherent rectangular bias of OBBs, we introduce a Mask-to-OBB loss, a differentiable formulation that enforces geometric consistency between predicted masks and OBB regions. Furthermore, we incorporate prompt-driven semantic guidance through two complementary modules-PAFE and DBFE-which enhance foreground representation and suppress background interference. Extensive experiments on 13 datasets across 5 imaging modalities show that OBBSeg not only outperforms existing weakly supervised methods but also achieves performance comparable to fully supervised approaches, demonstrating its potential for efficient and scalable medical image segmentation. The code is available at https://github.com/StarLxc3/OBBSeg.

Figures

Figures reproduced from arXiv: 2607.06007 by Chuhua Yang, Hui Huang, Jun Wei, Shuhui Wang, Xinchang Liu, Yu Liu.

Figure 1
Figure 1. Figure 1: Visualization and annotation cost (seconds per image) comparison among dif￾ferent annotation types (e.g., point, scribble, box, OBB, and mask). OBB annotation achieves a favorable balance between labeling efficiency and target coverage, providing geometry-aware and orientation-sensitive supervision with minimal annotation cost. with pixel-level annotations, but dense mask labeling is expensive and time￾con… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed OBBSeg framework. (a) Overall architecture. (b) The PAFE module enhances lesion-focused features via prompt guidance. (c) The DBFE module improves foreground–background discrimination. (d) Progressive mask gener￾ation from prompt inputs. supervision. Built upon a ViT [12] backbone, OBBSeg is organized into four stages. In each stage, the PAFE module exploits prompt cues to suppress… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed Mask-to-OBB (M2O) loss. The predicted mask s is projected into the OBB space to preserve spatial extent and orientation while sup￾pressing fine-grained shape details. The projected representation is aligned with the OBB label y, enabling geometry-consistent learning from OBB annotations. To obtain the prompt mask pi , we directly leverage user-provided prompts (e.g., points or scri… view at source ↗
Figure 4
Figure 4. Figure 4: Visualization comparison across three modalities. Competing Methods. We compare OBBSeg with 20 representative meth￾ods, including U-Net [34], U-Net++ [47], AttnUNet [29], TransUNet [6], PraNet [14], FLA-Net [25], MS-TFAL [11], SANet [42], LGRNet [44], MEGANet [5], CAS￾CADE [31], EMCAD [32], WeakPolyp [41], KnowSAM [18], BoxInst [38], Box￾Teacher [8], PointSup [7], AGMM [43], GazeMedSeg [46], and MedSAM2 [2… view at source ↗
Figure 5
Figure 5. Figure 5: (a) Impact of the proposed modules with the SAM2 backbone. (b) Visualization of representative failure cases and limitations [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of multi-organ segmentation on Synapse dataset. Multi-organ Segmentation. We evaluate OBBSeg on the 8-class Synapse dataset [17], achieving average Dice of 83.8%, 86.5%, 86.9%, 88.2% across Point, Box, Circle, Scribble prompts in [PITH_FULL_IMAGE:figures/full_fig_p012_6.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

46 extracted references · 46 canonical work pages · 2 internal anchors

  1. [1]

    Data in Brief28, 104863 (2020)

    Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in Brief28, 104863 (2020)

    work page 2020

  2. [2]

    In: ECCV

    Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What’s the point: Semantic segmentation with point supervision. In: ECCV. pp. 549–565. Springer (2016)

    work page 2016

  3. [3]

    saliency maps from physicians

    Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilar- iño, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. CMIG43, 99–111 (2015)

    work page 2015

  4. [4]

    The Cancer Imaging Archive (2015)

    Bloch, N., Madabhushi, A., Huisman, H., Freymann, J., Kirby, J., Grauer, M., Enquobahrie, A., Jaffe, C., Clarke, L., Farahani, K.: Nci-isbi 2013 challenge: Au- tomated segmentation of prostate structures. The Cancer Imaging Archive (2015). https://doi.org/10.7937/K9/TCIA.2015.ZF0VLOPV

    work page doi:10.7937/k9/tcia.2015.zf0vlopv 2013

  5. [5]

    In: WACV

    Bui, N.T., Hoang, D.H., Nguyen, Q.T., Tran, M.T., Le, N.: MEGANET: Multi- scale edge-guided attention network for weak boundary polyp segmentation. In: WACV. pp. 7985–7994 (2024)

    work page 2024

  6. [6]

    Chen, J., Mei, J., Li, X., Lu, Y., Yu, Q., Wei, Q., Luo, X., Xie, Y., Adeli, E., Wang, Y., et al.: Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers. MIA p. 103280 (2024)

    work page 2024

  7. [7]

    In: CVPR

    Cheng, B., Parkhi, O., Kirillov, A.: Pointly-Supervised Instance Segmentation. In: CVPR. pp. 2617–2626 (2022)

    work page 2022

  8. [8]

    In: CVPR

    Cheng, T., Wang, X., Chen, S., Zhang, Q., Liu, W.: Boxteacher: Exploring high- quality pseudo labels for weakly supervised instance segmentation. In: CVPR. pp. 3145–3154 (2023)

    work page 2023

  9. [10]

    In: ISBI

    Codella, N.C.F., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international sym- posium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: ISBI. pp. 168–...

    work page 2017

  10. [11]

    In: MICCAI

    Cui, B., Zhang, M., Xu, M., Wang, A., Yuan, W., Ren, H.: Rectifying noisy labels with sequential prior: Multi-scale temporal feature affinity learning for robust video segmentation. In: MICCAI. pp. 90–100. Springer (2023)

    work page 2023

  11. [12]

    In: ICLR (2021)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)

    work page 2021

  12. [13]

    IJCV88(2), 303–338 (2010)

    Everingham,M.,VanGool,L.,Williams,C.K.,Winn,J.,Zisserman,A.:Thepascal visual object classes (voc) challenge. IJCV88(2), 303–338 (2010)

    work page 2010

  13. [14]

    In: MICCAI

    Fan, D.P., Ji, G.P., Zhou, T., Chen, G., Fu, H., Shen, J., Shao, L.: Pranet: Par- allel reverse attention network for polyp segmentation. In: MICCAI. pp. 263–273. Springer (2020)

    work page 2020

  14. [15]

    TPAMI43(2), 652–662 (2019)

    Gao, S.H., Cheng, M.M., Zhao, K., Zhang, X.Y., Yang, M.H., Torr, P.: Res2net: A new multi-scale backbone architecture. TPAMI43(2), 652–662 (2019)

    work page 2019

  15. [16]

    In: ISBI

    Gong, H., Chen, G., Wang, R., Xie, X., Mao, M., Yu, Y., Chen, F., Li, G.: Multi- task learning for thyroid nodule segmentation with thyroid region prior. In: ISBI. pp. 257–261. IEEE (2021) OBBSeg 17

    work page 2021

  16. [17]

    org/10.7303/SYN3193805,https://repo- prod.prod.sagebase.org/repo/v1/ doi/locate?id=syn3193805&type=ENTITY

    harrigr: Segmentation outside the cranial vault challenge (2015).https://doi. org/10.7303/SYN3193805,https://repo- prod.prod.sagebase.org/repo/v1/ doi/locate?id=syn3193805&type=ENTITY

    work page doi:10.7303/syn3193805 2015

  17. [18]

    TMI (2025)

    Huang, K., Zhou, T., Fu, H., Zhang, Y., Zhou, Y., Gong, C., Liang, D.: Learnable prompting sam-induced knowledge distillation for semi-supervised medical image segmentation. TMI (2025)

    work page 2025

  18. [19]

    TMI33(2), 233–245 (2013)

    Jaeger, S., Karargyris, A., Candemir, S., Folio, L., Siegelman, J., Callaghan, F., Xue, Z., Palaniappan, K., Singh, R.K., Antani, S., et al.: Automatic tuberculosis screening using chest radiographs. TMI33(2), 233–245 (2013)

    work page 2013

  19. [20]

    In: International Con- ference on Multimedia Modeling

    Jha, D., Smedsrud, P.H., Riegler, M.A., Halvorsen, P., De Lange, T., Johansen, D., Johansen, H.D.: Kvasir-seg: A segmented polyp dataset. In: International Con- ference on Multimedia Modeling. pp. 451–462. Springer (2019)

    work page 2019

  20. [21]

    Machine Intelligence Research 19(6), 531–549 (2022)

    Ji, G.P., Xiao, G., Chou, Y.C., Fan, D.P., Zhao, K., Chen, G., Van Gool, L.: Video polyp segmentation: A deep learning perspective. Machine Intelligence Research 19(6), 531–549 (2022)

    work page 2022

  21. [22]

    In: ICCV

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV. pp. 4015–4026 (October 2023)

    work page 2023

  22. [23]

    MIA53, 165–178 (2019)

    Li, C., Wang, X., Liu, W., Latecki, L.J., Wang, B., Huang, J.: Weakly supervised mitosis detection in breast histopathology images using concentric loss. MIA53, 165–178 (2019)

    work page 2019

  23. [24]

    In: CVPR

    Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolu- tional networks for semantic segmentation. In: CVPR. pp. 3159–3167 (2016)

    work page 2016

  24. [25]

    In: MICCAI

    Lin, J., Dai, Q., Zhu, L., Fu, H., Wang, Q., Li, W., Rao, W., Huang, X., Wang, L.: Shifting more attention to breast lesion segmentation in ultrasound videos. In: MICCAI. pp. 497–507. Springer (2023)

    work page 2023

  25. [26]

    MedSAM2: Segment Anything in 3D Medical Images and Videos

    Ma, J., Yang, Z., Kim, S., Chen, B., Baharoon, M., Fallahpour, A., Asakereh, R., Lyu, H., Wang, B.: Medsam2: Segment anything in 3d medical images and videos. arXiv preprint arXiv:2504.03600 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [27]

    TMI34(10), 1993–2024 (2014)

    Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor image segmentation benchmark (brats). TMI34(10), 1993–2024 (2014)

    work page 1993

  27. [28]

    In: CVPR

    Oh, Y., Kim, B., Ham, B.: Background-aware pooling and noise-aware loss for weakly-supervised semantic segmentation. In: CVPR. pp. 6913–6922 (2021)

    work page 2021

  28. [29]

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

    Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., 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

  29. [30]

    In: International Symposium on Medical Information Processing and Analysis

    Pedraza, L., Vargas, C., Narváez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: International Symposium on Medical Information Processing and Analysis. vol. 9287, pp. 188–193. SPIE (2015)

    work page 2015

  30. [31]

    In: WACV

    Rahman, M.M., Marculescu, R.: Medical image segmentation via cascaded atten- tion decoding. In: WACV. pp. 6222–6231 (2023)

    work page 2023

  31. [32]

    In: CVPR

    Rahman, M.M., Munir, M., Marculescu, R.: EMCAD: Efficient multi-scale convo- lutional attention decoding for medical image segmentation. In: CVPR. pp. 11769– 11779 (2024)

    work page 2024

  32. [33]

    In: ICLR

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollar, P., Feichtenhofer, C.: Sam 2: Segment anything in images and videos. In: ICLR. vol. 2025, pp. 28085–28128 (2025)

    work page 2025

  33. [34]

    In: MICCAI

    Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomed- ical image segmentation. In: MICCAI. pp. 234–241. Springer (2015) 18 Wei et al

    work page 2015

  34. [35]

    In: ICML

    Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., et al.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: ICML. pp. 29441–29454. PMLR (2023)

    work page 2023

  35. [36]

    International Journal of Computer Assisted Radiology and Surgery9(2), 283–293 (2014)

    Silva, J., Histace, A., Romain, O., Dray, X., Granado, B.: Toward embedded detec- tion of polyps in wce images for early diagnosis of colorectal cancer. International Journal of Computer Assisted Radiology and Surgery9(2), 283–293 (2014)

    work page 2014

  36. [37]

    TMI35(2), 630–644 (2015)

    Tajbakhsh, N., Gurudu, S.R., Liang, J.: Automated polyp detection in colonoscopy videos using shape and context information. TMI35(2), 630–644 (2015)

    work page 2015

  37. [38]

    In: CVPR

    Tian, Z., Shen, C., Wang, X., Chen, H.: Boxinst: High-performance instance seg- mentation with box annotations. In: CVPR. pp. 5443–5452 (2021)

    work page 2021

  38. [39]

    Scientific Data5, 180161 (2018).https://doi.org/10.1038/sdata.2018.161

    Tschandl, P., Rosendahl, C., Kittler, H.: The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data5, 180161 (2018).https://doi.org/10.1038/sdata.2018.161

    work page doi:10.1038/sdata.2018.161 2018

  39. [40]

    Journal of Healthcare Engineering2017(1), 4037190 (2017)

    Vázquez, D., Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., López, A.M., Romero, A., Drozdzal, M., Courville, A.: A benchmark for endoluminal scene segmentation of colonoscopy images. Journal of Healthcare Engineering2017(1), 4037190 (2017)

    work page 2017

  40. [41]

    In: MICCAI

    Wei, J., Hu, Y., Cui, S., Zhou, S.K., Li, Z.: Weakpolyp: You only look bounding box for polyp segmentation. In: MICCAI. pp. 757–766. Springer (2023)

    work page 2023

  41. [42]

    In: MICCAI

    Wei, J., Hu, Y., Zhang, R., Li, Z., Zhou, S.K., Cui, S.: Shallow attention network for polyp segmentation. In: MICCAI. pp. 699–708. Springer (2021)

    work page 2021

  42. [43]

    In: CVPR

    Wu, L., Zhong, Z., Fang, L., He, X., Liu, Q., Ma, J., Chen, H.: Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In: CVPR. pp. 15454– 15464 (2023)

    work page 2023

  43. [44]

    In: MICCAI

    Xu, H., Yang, Y., Aviles-Rivero, A.I., Yang, G., Qin, J., Zhu, L.: Lgrnet: Local- global reciprocal network for uterine fibroid segmentation in ultrasound videos. In: MICCAI. pp. 667–677. Springer (2024)

    work page 2024

  44. [45]

    In: AAAI

    Zhang, X., Zhu, L., He, H., Jin, L., Lu, Y.: Scribble hides class: Promoting scribble- based weakly-supervised semantic segmentation with its class label. In: AAAI. pp. 7332–7340 (2024)

    work page 2024

  45. [46]

    In: MICCAI

    Zhong, Y., Tang, C., Yang, Y., Qi, R., Zhou, K., Gong, Y., Heng, P.A., Hsiao, J.H., Dou, Q.: Weakly-supervised medical image segmentation with gaze annotations. In: MICCAI. pp. 530–540. Springer (2024)

    work page 2024

  46. [47]

    TMI (2019)

    Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J.: Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. TMI (2019)

    work page 2019