Pith. sign in

REVIEW 5 major objections 5 minor 36 references

Robust Activation Map Rectification for Weakly Supervised Volumetric Segmentation: Temporal Coherence as a Free Lunch

T0 review · 5 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read A new training-free framework, CSSeg, claims that noisy class activation maps in weakly supervised volumetric segmentation can be fixed by exploiting temporal coherence, yielding up to 20% Dice and 40% mIoU gains with a fivefold inference s

desk verdict Integrated pipeline is a plausible practical contribution, but BER's detector as written fires on almost every frame, so the paper's central explanation for its main ablation gain does not hold. read the letter →

arxiv 2607.19877 v1 pith:SRT3IDUI submitted 2026-07-22 cs.CV

classification cs.CV
keywords weaklysupervisedsegmentationclassactivationmapsvolumetricmedicalimagestemporalcoherencevariancereductionrandommatrixtheoryfoundationmodelpromptingmaprectification
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

CSSeg claims that the main obstacle in weakly supervised volumetric segmentation is not learning better class activation maps (CAMs) but stabilizing them before they are used to prompt a large medical-image foundation model. The paper proposes a training-free, prototype-free pipeline: average CAMs across adjacent slices to cut noise variance, then detect slices whose activation support contradicts the aggregated map and replace them with a neighbor that scores higher on a tissue-density heuristic. After a bounding box is derived and a foundation model segments the slice, the same two steps are repeated once more. The authors report gains of up to 20.5% Dice and 40.3% mIoU over prior weakly supervised methods on public benchmarks, with roughly five times less inference time. The reason these results matter is that they suggest temporal coherence alone, without extra parameters or training stages, can serve as a strong inductive bias for medical volumes.

What carries the argument

The load-bearing objects are (1) Variance-Reduced Activation Aggregation (VRAA), a normalized average of slice-wise CAMs that concentrates the latent signal and shrinks noise as σ/√n; (2) Bidirectional Extremity Rectification (BER), an extremity check based on set inclusion between the frame mask and the aggregated mask (Eq. 10) that flags frames whose activation support is statistically inconsistent with the volume-level consensus, followed by conditional replacement using a brightness heuristic L(X,M) over the previous or next slice; and (3) a second pass in which the same VRAA+BER steps are reapplied to the foundation model's outputs (called DoubleSAM in the paper). The extremity check is

What would settle it

On a set of volumetric scans where the target is darker than its surroundings (hypointense on the relevant sequence), run the BER module with the brightness heuristic as given and with the comparison reversed. If the original heuristic picks the brighter but less accurate neighbor and Dice drops (or the reversed comparison yields higher Dice), the load-bearing assumption fails. Alternatively, a synthetic volume with a bright decoy region in the adjacent slice would settle it.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that noisy, unreliable CAMs in weakly supervised volumetric segmentation can be rectified by two parameter-free operations that exploit inter-slice coherence. VRAA aggregates the CAMs of all slices of a volume into a single activation map, reducing the variance of high-dimensional noise by a factor of 1/√n; theoretical justification is given by modeling each CAM as a latent signal plus zero-mean isotropic noise and invoking concentration of measure. BER then flags any frame whose binary activation mask is not contained in (or does not contain) the aggregated mask as statistically inconsistent, and replaces that frame's CAM and mask with the

Load-bearing premise

The method assumes that when a slice's activation map disagrees with the volume-level aggregated map, the neighboring slice whose mask covers a brighter region actually segments the target better, so the brightness heuristic L(X,M) is a reliable selector for replacing 'bad' frames.

Editorial extensions

If this is right

  • Because VRAA and BER are training-free and require only one hyperparameter (the threshold τ), they can be appended to any existing CAM-based weakly supervised pipeline, so the reported gains would stack on top of other CAM-refinement methods.
  • The variance-reduction argument implies the benefit should grow with slice count and activation-map dimensionality, making the approach more valuable for high-resolution volumes.
  • The dependence on adjacent-slice coherence means the method is best suited to smoothly varying anatomy; on data with abrupt slice-to-slice changes, the rectification may remove valid frames.
  • The more-than-fivefold inference speedup makes the approach feasible for clinical use where compute is limited.

Reading between the lines

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

  • A natural extension is to apply the same two-step rectification to other volumetric modalities (e.g., ultrasound or CT) or to video data with image-level labels, to test whether the coherence prior transfers beyond the medical benchmarks reported.
  • The brightness heuristic L(X,M) assumes higher density (brighter) tissue is the target; on datasets where lesions appear hypointense, replacing the raw brightness with a contrast score (inside vs outside the mask) would likely be more robust.
  • The random-matrix analysis assumes zero-mean, weakly dependent noise; an ablation that injects spatially correlated noise into CAMs would reveal how much of the gain depends on that idealization.
  • Because the method is model-agnostic on the segmentation side, swapping the large foundation model for a lightweight supervised segmenter would isolate whether the rectification, rather than the foundation model's capacity, drives the reported improvement.
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

5 major / 5 minor

Summary. The manuscript proposes CSSeg, a training-free post-processing framework for weakly supervised volumetric segmentation. Starting from class activation maps (CAMs) generated by an image classifier, the method (i) aggregates CAMs across slices (VRAA) to reduce noise, (ii) intersects each slice mask with the aggregate mask, (iii) applies Bidirectional Extremity Rectification (BER), which is intended to detect 'catastrophic' frames via the condition in Eq. (10) and replace them with adjacent frames when a brightness heuristic indicates higher tissue density, and (iv) feeds the resulting masks as bounding-box prompts to MedSAM, with the whole VRAA+BER procedure repeated a second time (DoubleSAM). Experiments on BraTS, CHAOS, and MSD report large improvements over weakly supervised and training-free baselines, with substantial inference-time reductions. The paper also provides a random-matrix-theory motivation for variance reduction and for the extremity check.

Significance. If the reported gains are reproducible, CSSeg would be a practically valuable, model-agnostic, training-free add-on to weakly supervised volumetric segmentation pipelines, requiring only one hyperparameter and exploiting temporal coherence without learning. The paper also makes a commendable attempt to ground the noise-reduction step in random matrix theory, and the VRAA variance calculation is correct under the stated i.i.d. assumption, with a reasonable extension to weak dependence in the appendix. However, the central theoretical justification for BER is logically flawed as written: the extremity condition in Eq. (10) is vacuous after the intersection step in Eq. (9), and the actual selection mechanism is an unvalidated brightness heuristic. The empirical evidence, while suggestive, is under-powered because all tables report single point estimates without error bars or significance tests, and no code is released. The idea is worth pursuing, but the present form does not adequately support the paper's strong claims.

major comments (5)
  1. [Sec. III-B.c and III-C.1, Eqs. (9)-(10)] The extremity condition is vacuous as written. Eq. (9) defines R(M_t) := R(M_t) ∩ R(S), so after this step R(M_t) ⊆ R(S). Consequently the first disjunct R(M_t) ⊈ R(S) in Eq. (10) is always false. The second disjunct R(S) ⊈ R(M_t) holds whenever the aggregate mask S contains any pixel outside the intersection, which is essentially every frame because S is formed by averaging over all slices. Thus Eq. (10) is satisfied for almost all frames, not selectively for 'catastrophic failures.' The accompanying text ('highest responses ... does not even fall within the region') is also inconsistent with the set-based formula. Please revise the detector so that it has real selectivity, or report the fraction of frames flagged and demonstrate that flagged frames are indeed failures.
  2. [Sec. III-C.3, Eqs. (14)-(15)] Because Eq. (10) is non-selective, the brightness comparison L(X_{t-1}, M_{t-1}) > L(X_t, M_t) is the effective decision rule. This heuristic is asserted without derivation or validation. The manuscript does not show that higher average intensity inside the mask correlates with better segmentation of the target. Consequently, the +0.093 Dice gain attributed to BER in Table IV cannot be taken as evidence for 'detecting implausible activations.' Please provide a quantitative validation (e.g., L vs. segmentation quality on a validation set) or an ablation that replaces the heuristic with a neutral rule and reports the outcome.
  3. [Sec. III-C.2, Eq. (13)] The random-matrix argument concerns the scale of noise maxima in continuous CAM heatmaps and states a separation condition A ≳ Cσ√(log d/n) that is assumed, not derived. It does not imply the binary-mask subset condition of Eq. (10). The sentence 'The extremity check works because ...' is an assumption, not a consequence of the preceding bounds. Please either prove a statement that relates signal dominance to the specific subset/inclusion criterion, or explicitly label the connection as an intuition/hypothesis.
  4. [Sec. III-B.b, Eqs. (2) and (8)] The paper states that 'the exact same τ' is used in Eq. (8), but Eq. (2) thresholds at τ · max(M_t) while Eq. (8) thresholds at (1−τ) · max(S). With τ = 0.8 these are very different thresholds (80% vs. 20% of the maximum). This inconsistency is important because it controls the size of the aggregate mask S and hence the behavior of BER. Please correct the typo or justify the asymmetry.
  5. [Sec. IV, Tables I-III and IV] All quantitative claims are single point estimates without standard deviations, confidence intervals, or significance tests. The comparisons with baseline methods (S2C, IPSeg, Clip-ES) also depend on how those methods are adapted and run, which is described only narratively. Given the abstract's strong claims ('up to 20% Dice and 40% mIoU gains'), please report variability across runs/folds and release code (or at least detailed configuration) so that the results can be reproduced. The current evidence is suggestive but insufficient for a journal-level claim.
minor comments (5)
  1. [Sec. IV-E, Fig. 8(b)] The text refers to 'after AMS and BER respectively,' but the method components are VRAA and BER; the abbreviation 'AMS' is undefined and appears to be a typo.
  2. [Notation throughout] The symbol M_t is used both for the continuous CAM (e.g., Eq. (3)) and for the binary mask (e.g., Eq. (2)). Please use distinct symbols (e.g., A_t for the heatmap and M_t for the binary mask) to avoid confusion, especially in Eqs. (9)-(15) and Fig. 3.
  3. [Fig. 6 caption] Typo: 'From let to right' should be 'From left to right.'
  4. [Sec. III-D] The DoubleSAM second pass is not motivated. Applying the same VRAA+BER procedure a second time to already-rectified masks is not obviously beneficial; please provide an explanation or an ablation varying the number of passes.
  5. [Appendix, Eq. (19)] The display of Eq. (19) is malformed: the norm symbol appears twice and the inequality is unclear. Please clean up the formatting.

Circularity Check

1 steps flagged · score 2.0 of 10

Central segmentation gains are empirical and not circular; BER's extremity check is, however, self-referential by construction.

  1. other [Section III-B.c and III-C.1 (Eqs. 9-10); replacement rule in Eqs. 14-15]
    "R(Mt) = R(Mt) ∩ R(S) ... R(Mt)⊈R(S)∨ R(S)⊈R(M t) ... S tends to cover a larger area than the segmentation mask of any individual frame."

    Eq. (9) defines R(Mt) to be the intersection with R(S), so R(Mt) ⊆ R(S) always. Hence the first disjunct of Eq. (10)'s 'extremity condition' is always false, and the second is true for every frame whose mask does not cover the whole aggregate S — i.e., essentially all frames, given the paper's own statement that S covers a larger area than individual frames. The 'catastrophic failure' detector thus reduces to a near-tautology; the actual selective decision is made by the brightness comparison L(X_{t-1},M_{t-1}) > L(X_t,M_t), which is asserted without derivation. The BER ablation gain is therefore attributed to a detection mechanism that is, by the paper's own equations, non-selective.

full rationale

The central claims of the paper — the Dice/mIoU improvements on BraTS, CHAOS, and MSD — are empirical comparisons against external benchmarks and do not reduce to a fitted parameter or to a self-citation chain. The VRAA variance-reduction statement is a direct consequence of the i.i.d. noise model in Eqs. (3)-(5) and is used only as motivation, not as a fitted prediction. The self-citations ([1], [24]) are related-work mentions and are not load-bearing for the derivation. The BER extremity check has a genuine self-referential problem: after Eq. (9), Eq. (10) is satisfied for almost every frame, so the claimed selective detection of 'catastrophic failures' is a formal artifact of the definitions rather than an empirical diagnosis; the actual selector is the brightness heuristic in Eqs. (14)-(15). This is a correctness/explanation weakness in one component, not a circular derivation of the main results. Accordingly, no central prediction is forced by construction, and the overall circularity score is low.

Assumptions & free parameters 1 free parameters · 7 assumptions · 0 invented entities

The method depends on a single hand-set threshold τ, a noise model that is assumed rather than checked, a coherence assumption for anatomy, and two ad hoc heuristics (the non-subset detector and brightness-based replacement). No new physical entities are introduced.

free parameters (1)
  • tau (CAM threshold) = 0.8 (set for all benchmarks)
    Threshold used in Eq. (2) and Eq. (8) to binarize CAMs and the aggregated map. The paper states performance is insensitive for τ in [0.5, 0.8], but the value is chosen by hand and the central results depend on it.
assumptions (7)
  • domain assumption CAMs are latent signal plus zero-mean isotropic sub-Gaussian noise (Eq. 3)
    Section III-A: the entire VRAA justification rests on modeling M_t = M*_t + σ ξ_t with zero-mean isotropic noise. This is asserted, not validated on real CAM distributions.
  • domain assumption The latent signal M*_t is approximately constant across slices; adjacent frames are coherent
    Section III-B and Section IV-E ('multi-frame consistency is desired, adjacent frames shall not change abruptly') assume that averaging over slices preserves the true target while canceling noise. This is often violated for organs/tumors that change shape and position.
  • standard math Noise vectors across slices are independent or weakly dependent with summable autocovariance (Eq. 21)
    The Appendix extends the variance calculation to correlated noise under the summability condition Σ||T_k|| < ∞. This is a standard sufficient condition but is not verified for the datasets.
  • ad hoc to paper The non-subset condition in Eq. (10) identifies frames with implausible activations
    Section III-C: the paper claims the condition detects 'statistical inconsistency,' but for any two distinct binary masks at least one non-subset relation holds, so the detector is not selective as stated.
  • ad hoc to paper Average brightness inside the mask (L(X,M)) indicates tissue density and is comparable across adjacent frames
    Eqs. (14)-(15): the replacement rule assumes that a brighter neighbor mask is a better segmentation. This is a heuristic with no supporting evidence in the paper.
  • domain assumption Bounding boxes extracted from CAM-derived masks are suitable prompts for MedSAM
    Section III-D cites [30] for improved bounding boxes leading to better segmentation. This is reasonable but external, and the paper does not study prompt sensitivity.
  • ad hoc to paper Running the same VRAA+BER pipeline a second time (DoubleSAM) further improves accuracy
    Section III-D and Table IV: no derivation is given for why a second pass should help; the ablation shows a modest gain, but the mechanism is unexplained.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Activation Map Rectification for Weakly Supervised Volumetric Segmentation: Temporal Coherence as a Free Lunch." pith.science (2026). https://pith.science/paper/SRT3IDUI

@misc{pith2026260719877,
  author       = {Pith},
  title        = {Pith review of: Robust Activation Map Rectification for Weakly Supervised Volumetric Segmentation: Temporal Coherence as a Free Lunch},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRT3IDUI}},
  note         = {Machine review of arXiv:2607.19877}
}
read the original abstract

Weakly supervised segmentation relies heavily on class activation maps (CAMs) to initially localize target regions. However, CAMs are often noisy and prone to catastrophic failures. Existing remedies typically introduce additional training stages or prototype learning, increasing computational cost and reducing robustness. In this paper, we propose a training-free prototype-free framework that rectifies unreliable CAMs by exploiting temporal and structural coherence in volumetric data as a free lunch. Our approach is built on two key components. First, we introduce Variance-Reduced Activation Aggregation (VRAA) which suppresses noise and amplify coherent semantic signals. We provide a theoretical justification by modeling CAMs as high-dimensional random vectors and show that aggregation yields provable variance reduction. Second, we design a Bidirectional Extremity Rectification (BER) mechanism that detects and rectifies implausible activations through bidirectional extremity checks, effectively mitigating extreme-value failures without learning additional parameters. Our method is model-agnostic and can be seamlessly integrated with existing pipelines. Extensive experiments on multiple public benchmarks demonstrate substantial improvements over state-of-the-art weakly supervised methods, achieving up to 20% Dice and 40% mIoU gains while reducing inference time by more than 5 times. These results indicate that leveraging coherence as an implicit inductive bias yields a principled and efficient approach to stabilizing weakly supervised volumetric segmentation. Our code will be available.

Figures

Figures reproduced from arXiv: 2607.19877 by the authors.

Figure 1
Figure 1. Different segmentation paradigms using large foundation models (LFMs). (a) Fully supervised methods. (b) Weakly supervised segmentation frameworks that transfer the knowledge of LFM to the classifier by training, such as S2C [20], FMA-WSSS [21], and CG-CDM [23]. (c) Open-world segmentation, such as IPSeg [22], with a prototype image prompt. (d) Our prototype-free weakly supervised segmentation requires no training o… view at source ↗
Figure 2
Figure 2. In Section III-A, the initial CAMs are first acquired. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method. (A) Our overall workflow. Our framework includes the Variance-Reduced Activation Aggregation (VRAA strategy, the Bidirectional Extremity Rectification (BER) module, and postprocessing (Sec. III-D). (B) Details of the VRAA strategy (Sec. III-B). (C) The BER module. Further details can be found in Sec. III-C. (D) The MedSAM-based segmentation with postprocessing (Sec. III-D). in uncert… view at source ↗
Figures from the paper (7 more)
Figure 3
Figure 3. Figure 3: Bidirectional Extremity check. The top row checks if the highest responses of the aggregated activation map fall out of the region of frame t’s binary mask Mt. The bottom row checks if the highest responses of CAM at frame t fall out of the region of the binary mask S …
Figure 4
Figure 4. Figure 4: Segmentation performance on the BraTS dataset. Red color highlights the segmented area. From left to right: input image; input CAM; Baseline (feeding the initial CAMs to MedSAM.); Baseline + VRAA (Sec. III-B); Baseline + VRAA + BER (Sec. III-C); Baseline + VRAA+ BER + …
Figure 5
Figure 5. Figure 5: Qualitative evaluation on CHAOS. From left to right: input image; initial CAM; baseline; CSSeg result; ground truth. CSSeg corrects noisy initial CAMs, yielding improved segmentation results on this challenging dataset. 3) Results on MSD: We further compare our method …
Figure 6
Figure 6. Figure 6: Qualitative evaluation on MSD. From let to right: input image; initial CAM; baseline; CSSeg result; ground truth. Our method significantly outperforms the baseline across MSD-Brain (top 2 rows), MSD-Prostate (middle 2 rows), and MSD-Cardiac (bottom 2 rows). TABLE IV AB…
Figure 7
Figure 7. Figure 7: Ablation study on hyperparameter τ [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Robustness to inaccurate initial CAMs. (a) We sort the 3D samples by initial CAM quality. CSSeg improves Dice over initial CAM consistently. (b) Distribution of adjacent-frame coherence is closer to GT after VRAA and BER [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Failure cases. When the target is too small, CSSeg is prone to failure. each activation map is an isotropic i.i.d. random vector. Below are standard ways to keep essentially the same kind of upper bound when ξt is not i.i.d. (e.g., temporally correlated across slides).…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 4 linked inside Pith

  1. [1]

    Segmamba-v2: Long-range sequential modeling mamba for general 3d medical image segmentation,

    Z. Xing, T. Ye, Y . Yang, D. Cai, B. Gai, X.-J. Wu, F. Gao, and L. Zhu, “Segmamba-v2: Long-range sequential modeling mamba for general 3d medical image segmentation,”IEEE Transactions on Medical Imaging, 2025

  2. [2]

    Cqformer: Learning dynamics across slices in medical image segmentation,

    S. Zhang, X. Shen, X. Chen, Z. Yu, B. Ren, H. Yang, X.-Y . Zhang, and Y . Zhou, “Cqformer: Learning dynamics across slices in medical image segmentation,”IEEE Transactions on Medical Imaging, vol. 44, no. 2, pp. 1043–1057, 2024

  3. [3]

    Gm-abs: promptable generalist model drives active barely supervised training in specialist model for 3d medical image segmentation,

    Z. Xu, C. Chen, D. Lu, J. Sun, D. Wei, Y . Zheng, Q. Li, and R. K.-y. Tong, “Gm-abs: promptable generalist model drives active barely supervised training in specialist model for 3d medical image segmentation,”IEEE Transactions on Medical Imaging, 2025

  4. [4]

    Metassl: A general heterogeneous loss for semi-supervised medical image segmentation,

    W. Zhao, L. Zhong, X. Liao, W. Liao, S. Zhang, S. Zhang, and G. Wang, “Metassl: A general heterogeneous loss for semi-supervised medical image segmentation,”IEEE Transactions on Medical Imaging, vol. 45, no. 2, pp. 751–763, 2026

  5. [5]

    Unlocking the potential of ordinary classifier: Class-specific adversarial erasing frame- work for weakly supervised semantic segmentation,

    H. Kweon, S.-H. Yoon, H. Kim, D. Park, and K.-J. Yoon, “Unlocking the potential of ordinary classifier: Class-specific adversarial erasing frame- work for weakly supervised semantic segmentation,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 6994–7003

  6. [6]

    Anti-adversarially manipulated attributions for weakly and semi-supervised semantic segmentation,

    J. Lee, E. Kim, and S. Yoon, “Anti-adversarially manipulated attributions for weakly and semi-supervised semantic segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 4071–4080

  7. [7]

    Railroad is not a train: Saliency as pseudo-pixel supervision for weakly supervised semantic segmentation,

    S. Lee, M. Lee, J. Lee, and H. Shim, “Railroad is not a train: Saliency as pseudo-pixel supervision for weakly supervised semantic segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 5495–5505

  8. [8]

    Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation,

    L. Xu, W. Ouyang, M. Bennamoun, F. Boussaid, F. Sohel, and D. Xu, “Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 6984–6993. 10 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. XX, NO. XX, XXXX 2026

Show all 36 references
  1. [9]

    Weakly supervised semantic segmentation by pixel-to-prototype contrast,

    Y . Du, Z. Fu, Q. Liu, and Y . Wang, “Weakly supervised semantic segmentation by pixel-to-prototype contrast,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4320–4329

  2. [10]

    Regional semantic contrast and aggregation for weakly supervised semantic segmentation,

    T. Zhou, M. Zhang, F. Zhao, and J. Li, “Regional semantic contrast and aggregation for weakly supervised semantic segmentation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4299–4309

  3. [11]

    Conditional diffusion models for weakly supervised medical image segmentation,

    X. Hu, Y .-J. Chen, T.-Y . Ho, and Y . Shi, “Conditional diffusion models for weakly supervised medical image segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Inter- vention. Springer, 2023, pp. 756–765

  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 . Loet al., “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4015–4026

  5. [13]

    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, p. 654, 2024

  6. [14]

    Semantic-aware sam for point-prompted instance segmentation,

    Z. Wei, P. Chen, X. Yu, G. Li, J. Jiao, and Z. Han, “Semantic-aware sam for point-prompted instance segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3585–3594

  7. [15]

    Asam: Boosting segment anything model with adversarial tuning,

    B. Li, H. Xiao, and L. Tang, “Asam: Boosting segment anything model with adversarial tuning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3699–3710

  8. [16]

    Eviprompt: A training-free evidential prompt generation method for adapting segment anything model in medical images,

    Y . Xu, J. Tang, A. Men, and Q. Chen, “Eviprompt: A training-free evidential prompt generation method for adapting segment anything model in medical images,”IEEE Transactions on Image Processing, 2024

  9. [17]

    Alps: An auto-labeling and pre-training scheme for remote sensing segmentation with segment anything model,

    S. Zhang, Q. Wang, J. Liu, and H. Xiong, “Alps: An auto-labeling and pre-training scheme for remote sensing segmentation with segment anything model,”IEEE Transactions on Image Processing, 2025

  10. [18]

    Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation,

    Y . Lin, M. Chen, W. Wang, B. Wu, K. Li, B. Lin, H. Liu, and X. He, “Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023,...

  11. [19]

    Clims: Cross language image matching for weakly supervised semantic segmentation,

    J. Xie, X. Hou, K. Ye, and L. Shen, “Clims: Cross language image matching for weakly supervised semantic segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 4483–4492

  12. [20]

    From sam to cams: Exploring segment anything model for weakly supervised semantic segmentation,

    H. Kweon and K.-J. Yoon, “From sam to cams: Exploring segment anything model for weakly supervised semantic segmentation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 499–19 509

  13. [21]

    Foundation model assisted weakly supervised semantic segmentation,

    X. Yang and X. Gong, “Foundation model assisted weakly supervised semantic segmentation,” inProceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision, 2024, pp. 523–532

  14. [22]

    Towards training-free open- world segmentation via image prompt foundation models,

    L. Tang, P.-T. Jiang, H. Xiao, and B. Li, “Towards training-free open- world segmentation via image prompt foundation models,”International Journal of Computer Vision, vol. 133, no. 1, pp. 1–15, 2025

  15. [23]

    Conditional diffusion models for weakly supervised medical image segmentation,

    X. Hu, Y .-J. Chen, T.-Y . Ho, and Y . Shi, “Conditional diffusion models for weakly supervised medical image segmentation,” inMedical Im- age Computing and Computer Assisted Intervention – MICCAI 2023, H. Greenspan, A. Madabhushi, P. Mousavi, S. Salcudean, J. Duncan, T. Syeda...

  16. [24]

    Medical sam 2: Segment medical images as video via segment anything model 2,

    J. Zhu, Y . Qi, and J. Wu, “Medical sam 2: Segment medical images as video via segment anything model 2,”arXiv preprint arXiv:2408.00874, 2024

  17. [25]

    Learning pixel-level semantic affinity with image- level supervision for weakly supervised semantic segmentation,

    J. Ahn and S. Kwak, “Learning pixel-level semantic affinity with image- level supervision for weakly supervised semantic segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4981–4990

  18. [26]

    Cian: Cross-image affinity net for weakly supervised semantic segmentation,

    J. Fan, Z. Zhang, T. Tan, C. Song, and J. Xiao, “Cian: Cross-image affinity net for weakly supervised semantic segmentation,” inProceed- ings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 10 762–10 769

  19. [27]

    An image is worth 16x16 words: Transformers for image recognition at scale,

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

  20. [28]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 618–626

  21. [29]

    Vershynin,High-dimensional probability: An introduction with ap- plications in data science

    R. Vershynin,High-dimensional probability: An introduction with ap- plications in data science. Cambridge university press, 2018, vol. 47

  22. [30]

    Sam- med2d,

    J. Cheng, J. Ye, Z. Deng, J. Chen, T. Li, H. Wang, Y . Su, Z. Huang, J. Chen, L. J. H. Sun, J. He, S. Zhang, M. Zhu, and Y . Qiao, “Sam- med2d,” 2023

  23. [31]

    The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification,

    U. Baid, S. Ghodasara, S. Mohan, M. Bilello, E. Calabrese, E. Colak, K. Farahani, J. Kalpathy-Cramer, F. C. Kitamura, S. Patiet al., “The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification,”arXiv preprint arXiv:2107.02314, 2021

  24. [32]

    Chaos challenge- combined (ct-mr) healthy abdominal organ segmentation,

    A. E. Kavur, N. S. Gezer, M. Barıs ¸, S. Aslan, P.-H. Conze, V . Groza, D. D. Pham, S. Chatterjee, P. Ernst, S. ¨Ozkanet al., “Chaos challenge- combined (ct-mr) healthy abdominal organ segmentation,”Medical Image Analysis, vol. 69, p. 101950, 2021

  25. [33]

    The medical segmentation decathlon,

    M. Antonelli, A. Reinke, S. Bakas, K. Farahani, A. Kopp-Schneider, B. A. Landman, G. Litjens, B. Menze, O. Ronneberger, R. M. Summers et al., “The medical segmentation decathlon,”Nature communications, vol. 13, no. 1, p. 4128, 2022

  26. [34]

    Diffusion models for medical anomaly detection,

    J. Wolleb, F. Bieder, R. Sandk ¨uhler, and P. C. Cattin, “Diffusion models for medical anomaly detection,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:247318844

  27. [35]

    Sam 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huang, J. Lei, T. Ma, B. Guo, A. Kalla, M. Marks, J. Greer, M. Wang, P. Sun, R. R ¨adle, T. Afouras, E. Mavroudi, K. Xu, T.-H. Wu, Y . Zhou, L. Momeni, R. Hazra, S. Ding, S. ...

  28. [36]

    Medsam2: Segment anything in 3d medical images and videos,

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

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.