Pith. sign in

REVIEW 3 major objections 4 minor 42 references

ScSAM: Debiasing Morphology and Distributional Variability in Subcellular Semantic Segmentation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ScSAM fuses SAM with MAE embeddings to set top scores in subcellular segmentation.

desk verdict Competent fusion architecture for EM subcellular segmentation, but the state-of-the-art claims are currently undermined by MAE pre-training on the evaluation set. read the letter →

arxiv 2507.17149 v1 pith:T7GFOVVI submitted 2025-07-23 cs.CV cs.AIcs.LG

Bo Fang , Jianan Fan , Dongnan Liu , Hang Chang , Gerald J.Shami , Filip Braet , Weidong Cai This is my paper
classification cs.CVcs.AIcs.LG
keywords subcellularsegmentationelectronmicroscopySegmentAnythingModelMaskedAutoencoderfeaturefusionclasspromptencodercross-taskalignmentorganelle
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

The paper claims that subcellular organelle segmentation in electron microscopy is improved by fusing two frozen foundation-model encoders instead of relying on one. Its method, ScSAM, pairs the Segment Anything Model with a Masked Autoencoder whose embeddings act as a cellular prior, aligning both in a shared feature space and activating class-specific regions with a learned prompt encoder. On the BetaSeg islet-cell data, ScSAM reports the top score on every aggregate metric in both high- and low-glucose settings, with mean IoU up 11.3% in low-glucose. The paper's message is that the bias caused by organelle shape and distribution variability is largely addressable at the representation level, without fine-tuning the backbones.

What carries the argument

The load-bearing component is the Feature Alignment and Fusion Module: it projects frozen SAM and MAE embeddings into a common dimension with small MLPs, aligns their directions with a cosine-similarity loss, concatenates and convolves them with Group Normalization, re-weights channels with a channel attention module, and adds an auxiliary SAM-only stream to preserve semantic detail. The second load-bearing piece is the cosine-similarity class prompt encoder: learnable class prototypes are compared with the fused feature map, the resulting similarity matrix activates class-specific regions, and one-hot coding turns the activated features into dense and sparse prompts for the mask decoder. A contrastive loss pulls each prototype toward its class embeddings and pushes different classes apart, while a Dice loss supervises the final masks.

What would settle it

Run the same ScSAM pipeline with an MAE encoder pre-trained on unrelated images, such as natural photographs or a different electron-microscopy dataset, and evaluate on the same BetaSeg splits; if the margin over the SAM-only baseline mostly disappears, the reported gains come from target-domain familiarity in the MAE rather than from the fusion mechanism itself.

Watch

Extended reading notes

Core claim

ScSAM's central claim is that SAM's failure modes on subcellular data, confusing similarly shaped small organelles and missing fine detail, come from relying on a single feature mapping, and that the cure is a complementary mapping supplied by a Masked Autoencoder. The paper shows that MAE embeddings, learned by image reconstruction, attend to multi-scale structural patterns, while SAM embeddings encode edges, shapes, and region-level consistency. The proposed method aligns these two spaces with a cosine-similarity loss, fuses them through attention-weighted convolutions, and feeds the fused features together with class prototypes to the mask decoder. The paper reports the top Challenge IoU, mean IoU, AJI, and overall Dice score in both nutritional settings of BetaSeg, with the largest gains on mitochondria and granules.

Load-bearing premise

The argument rests on treating the Masked Autoencoder's weights, which were pre-trained on the same BetaSeg image volumes used for evaluation, as generic cellular prior knowledge rather than as advance familiarity with the test images.

Editorial extensions

If this is right

  • The dual frozen-encoder design means the same SAM and MAE pair can be attached to other segmentation heads while updating only a small set of parameters, reducing the cost of adapting foundation models to microscopy.
  • The largest reported gains are on small, irregular organelles: granule Dice rises from 0.505 with vanilla SAM to 0.798 in high-glucose cells, and from 0.385 to 0.767 in low-glucose cells.
  • Training reaches peak Dice in about 3.2 hours, roughly half the training time reported for the SurgicalSAM baseline under the same settings.
  • Cross-dataset tests in both glucose-transfer directions show the fused representation beats a SAM-only baseline, with up to 5.3% absolute Dice improvement on granules.

Reading between the lines

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

  • Beyond the paper: because the MAE encoder was pre-trained on the same BetaSeg volumes used for evaluation, the cleanest control would be replacing it with an MAE trained on unrelated images; the paper does not include that control.
  • Beyond the paper: the same cross-task fusion could be tested on volume electron microscopy and other long-tailed biomedical segmentation tasks where one encoder supplies texture and the other supplies boundaries.
  • Beyond the paper: the class prompt encoder's prototype-similarity activation resembles a few-shot mechanism, so a natural extension is segmenting a novel organelle from a single annotated prototype without retraining.
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

3 major / 4 minor

Summary. The paper proposes ScSAM, a subcellular semantic segmentation framework that fuses a frozen SAM encoder with a frozen MAE encoder. A Feature Alignment and Fusion Module (FAFM) aligns and combines the two embedding spaces, and a cosine-similarity-based class prompt encoder generates class-specific sparse and dense prompts for the mask decoder. The method is evaluated on the BetaSeg high- and low-glucose FIB-SEM datasets, where it reports state-of-the-art results on Challenge IoU, m IoU, AJI, and Dice, with ablations attributing gains to the fusion module, alignment loss, and class prompt encoder.

Significance. If the reported gains survive a clean evaluation, ScSAM provides a practical recipe for adapting SAM to subcellular EM under limited labels, and it is one of the first SAM-based approaches specifically targeting organelle segmentation. The paper includes thorough ablations, efficiency comparisons, and qualitative visualizations. However, the target-domain MAE pre-training and the absence of error bars currently make the headline claims conditional. The central architecture is sensible, but the empirical support for generalizable 'cellular prior knowledge' is not yet established.

major comments (3)
  1. [Supplementary 1.1; Section 4.3] The MAE encoder used as the 'cellular prior' is pre-trained on the BetaSeg datasets themselves, with no reported split between pre-training and evaluation (Supplementary 1.1). The main text presents this as general prior knowledge (Section 3.2), but the evaluation in Tables 1, 2, and 4 compares a target-adapted encoder plus SAM against SAM-only baselines. Because the MAE was pre-trained on both high- and low-glucose BetaSeg data, the 'cross-dataset' experiment in Table 4 does not test cross-domain transfer in the encoder. Please add a variant with MAE pre-trained only on the training-split cells (or on unrelated cellular EM data) and re-report the same tables. Without this, the 11.3% mIoU gain cannot be attributed to the fusion mechanism rather than to target-domain self-supervised pre-training.
  2. [Section 4.1; Tables 1 and 2] The experimental protocol uses a single islet cell per dataset for training and the remaining cells for validation, and all reported metrics are single-run values without error bars or significance tests. Given the small validation set and the high variance typical of EM segmentation, the claim that 'ScSAM delivers the top score on every aggregate metric' is not yet supported with appropriate uncertainty quantification. Please report mean±std over multiple seeds or leave-one-cell-out cross-validation, and indicate whether the margins (e.g., the 11.3% mIoU improvement in low-glucose) are statistically significant.
  3. [Supplementary 1.2; Tables 1 and 2] The evaluation protocol for Vanilla SAM and SAM2 assigns predicted contours to organelle categories based on overlap with ground truth or nearest-contour mapping. This is not a standard prompt-free semantic segmentation protocol and can systematically favor or disfavor these baselines relative to the other methods, which are trained end-to-end. Please either justify this protocol with reference to standard practice or re-evaluate SAM/SAM2 under the same protocol used for the other frozen-encoder baselines, so the comparison with ScSAM is fair.
minor comments (4)
  1. [Section 4.4] The sentence 'Fig. 3 visualizes the pixel-level positive class similarity maps' should refer to Figure 6, not Figure 3, which shows the architecture.
  2. [Table 3 (main text)] Rows in Table 3 use only three checkmarks each, making it ambiguous which component is removed; please use explicit '✓'/'✗' for all four components (Fuse, Lcos, Dens, Spar).
  3. [Figure 3 and Figure 4] Several abbreviations (CC, GN, CAM, NTX) are used in the figures but are not defined in the captions; please define them or refer the reader to the relevant equations.
  4. [Overall manuscript] The fact that the MAE is pre-trained on the BetaSeg datasets is currently only mentioned in the supplementary material; please state this explicitly in the main text, as it is critical for interpreting all experimental results.

Circularity Check

2 steps flagged · score 6.0 of 10

MAE pretrained on the same BetaSeg datasets used for evaluation makes the claimed 'cellular prior' and cross-dataset generalization results non-independent.

  1. fitted input called prediction [Supplementary Section 1.1 ('Pre-trained MAE'); Section 4.1 ('Datasets and Evaluation'); Section 4.2 ('Implementation Details')]
    "MAE is pre-trained on the BetaSeg datasets, utilizing a cropped image size of 256, a patch size of 16, a mask ratio of 0.5, and 1600 epochs. ... the last cell of the two datasets serves as the training set, and ScSAM is validated on the remaining cells."

    The claimed 'MAE-guided cellular prior knowledge' is the MAE encoder, and Supplementary 1.1 states that it is pre-trained on 'the BetaSeg datasets'. Section 4.1 then defines the evaluation split on those same datasets, with one cell as training and the remaining cells as validation. On the paper's own description, the pretraining corpus is the same BetaSeg datasets later divided into training and validation cells; no statement excludes the validation cells from the 1600-epoch MAE pretraining. The frozen MAE encoder has therefore already been fitted to the images on which Tables 1 and 2 report 'predictions'.

  2. fitted input called prediction [Section 4.6 ('Cross-Dataset Generalization'), Table 4; Supplementary Section 1.1]
    "MAE is pre-trained on the BetaSeg datasets. ... We further evaluate ScSAM's robustness and transferability for subcellular recognition by training on one dataset and testing on another. ... T, V, H, and L represent the training and validation sets and the high- and low-glucose BetaSeg datasets, respectively."

    The 'cross-dataset generalization' experiment is described as transferring between high- and low-glucose BetaSeg, but the MAE pretraining used 'the BetaSeg datasets', meaning both glucose subsets. In both rows of Table 4, the validation dataset (low-glucose in row 1, high-glucose in row 2) is contained in the MAE pretraining corpus. The claimed 'absolute Dice improvement of 5.3% for granule segmentation' therefore cannot be attributed to domain-invariant cross-task fusion; the MAE has already reconstructed images from the validation domain. This particular 'generalization' prediction reduces, by construction of the pretraining corpus, to an in-distribution evaluation of a target-pretrained encoder, not a test on an unseen dataset.

full rationale

There is no equation-level circularity in the proposed modules: FAFM, the cosine class prompt encoder, and the Dice/NTXent losses are external design choices, and the paper ablates them against concatenation, cross-attention, and single-encoder baselines. The circularity burden is at the data level and is load-bearing for the main empirical claim. Supplementary Section 1.1 states that the MAE was pre-trained on 'the BetaSeg datasets' for 1600 epochs, while Section 4.1 defines the segmentation training and validation splits from those same BetaSeg datasets. With no reported split for the MAE pretraining corpus, the frozen MAE encoder has been fitted to the target-domain images before ScSAM's downstream training. The main tables and especially Section 4.6's cross-dataset experiment therefore compare a target-pretrained encoder plus SAM against SAM-only baselines, so the reported 11.3% mIoU gain and 5.3% granule Dice gain may reflect target-domain self-supervised pretraining rather than the proposed fusion or prompt mechanism. The near-random Dice of MAE* does not refute this, because raw MAE embeddings are not directly segmentation-ready yet can still inject target-domain statistics after FAFM alignment and class-prompt training. This is partial circularity: the fitted input (a MAE trained on the evaluation domain) is presented as a generalizable prior, and the 'prediction' is measured on that same domain. The independent content of FAFM and the class prompt encoder prevents a higher score, but the central claim's empirical support is compromised.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method rests on empirical assumptions about transferable representations and data representativeness; no new physical entities are proposed. The class prototype bank and learned prompts are model parameters, not external postulates. The main tuning parameters are the cosine-loss weight, the NT-Xent temperature, prototype dimensions, and the MAE pre-training recipe. The most consequential assumption is that MAE pre-training on the evaluation dataset is a fair prior rather than target-domain leakage.

free parameters (4)
  • lambda, weight of cosine alignment loss = 0.2 (chosen from 0.1, 0.2, 0.3 in Supplementary Table 1)
    Total loss in Eq. 9 is L = lambda*Lcos + L_NTX + L_Dice; the paper tunes lambda on BetaSeg, so reported performance depends on this validation-set choice.
  • tau, temperature in NT-Xent loss = not reported
    Eq. 8 includes a temperature parameter tau; the implementation details section does not state its value, so replication requires guessing or using a default.
  • class prototype dimension and mask decoder hidden units = 256 and 128
    Section 4.2 sets these dimensions by hand; they control capacity of the class prompt encoder and decoder but are not derived from theory.
  • MAE pre-training mask ratio and duration = mask ratio 0.5, 1600 epochs
    Supplementary 1.1 fixes the MAE pre-training recipe by hand; because pre-training uses the target-domain BetaSeg images, this choice is a confound for the prior-knowledge claim.
assumptions (5)
  • domain assumption Frozen SAM and MAE embeddings provide complementary information for subcellular images.
    The method assumes the two encoders encode orthogonal cues (edges/global structure versus texture/multi-scale patterns); evidence is qualitative activation maps in Fig. 2, not a quantitative guarantee.
  • domain assumption MAE pre-trained on BetaSeg images provides useful cellular prior knowledge.
    Supplementary 1.1 states MAE is pre-trained on BetaSeg (target-domain) datasets; the paper does not evaluate a version pre-trained on unrelated data, so the transfer claim is untested.
  • domain assumption Aligning embeddings by cosine similarity in a shared space preserves discriminative information needed by the decoder.
    FAFM design assumes minimizing Lcos does not destroy class-separating structure; ablations support this empirically, but there is no theoretical guarantee.
  • domain assumption The selected middle 350 slices and last-cell train split represent the distribution of subcellular structures.
    Section 4.1 selects slices and splits by cell; there is no evidence that this subset is representative or that single-cell training supports generalizable conclusions.
  • domain assumption NT-Xent contrastive loss on class prototypes improves intra-class compactness and inter-class separation.
    This is a standard contrastive-learning assumption; it is applied in Eq. 8 and validated only through the final segmentation metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ScSAM: Debiasing Morphology and Distributional Variability in Subcellular Semantic Segmentation." pith.science (2026). https://pith.science/paper/T7GFOVVI

@misc{pith2026250717149,
  author       = {Pith},
  title        = {Pith review of: ScSAM: Debiasing Morphology and Distributional Variability in Subcellular Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T7GFOVVI}},
  note         = {Machine review of arXiv:2507.17149}
}
read the original abstract

The significant morphological and distributional variability among subcellular components poses a long-standing challenge for learning-based organelle segmentation models, significantly increasing the risk of biased feature learning. Existing methods often rely on single mapping relationships, overlooking feature diversity and thereby inducing biased training. Although the Segment Anything Model (SAM) provides rich feature representations, its application to subcellular scenarios is hindered by two key challenges: (1) The variability in subcellular morphology and distribution creates gaps in the label space, leading the model to learn spurious or biased features. (2) SAM focuses on global contextual understanding and often ignores fine-grained spatial details, making it challenging to capture subtle structural alterations and cope with skewed data distributions. To address these challenges, we introduce ScSAM, a method that enhances feature robustness by fusing pre-trained SAM with Masked Autoencoder (MAE)-guided cellular prior knowledge to alleviate training bias from data imbalance. Specifically, we design a feature alignment and fusion module to align pre-trained embeddings to the same feature space and efficiently combine different representations. Moreover, we present a cosine similarity matrix-based class prompt encoder to activate class-specific features to recognize subcellular categories. Extensive experiments on diverse subcellular image datasets demonstrate that ScSAM outperforms state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2507.17149 by the authors.

Figure 1
Figure 1. Comparison of ScSAM against existing feature fusion segmentation models. Traditional methods need to update the parameters of backbones and require additional relevant datasets, whereas SAM-based fusion algorithms contain frozen backbones pre-trained for the same task. The proposed ScSAM fuses cross-task feature representations to problematic learning patterns in subcellular segmentation. tracking, can improve train… view at source ↗
Figure 2
Figure 2. Visualization of subcellular distribution and feature activation maps: (a) Masked frame, (b) and (c) are feature activation maps for SAM and MAE embeddings indicating complementary feature representations. The red response region is the region of interest of dual backbones, demonstrating orthogonal information dimensions in the embedders. handle intricate subcellular morphologies and subtle features [7, 40]. Recent … view at source ↗
Figure 3
Figure 3. The overview of ScSAM, which utilizes pre-trained MAE as prior knowledge related to subcellular for frozen SAM to enhance feature perception of organelle morphology and distribution in EMIs. We design the FAFM and cosine similarity-based class prompt encoder to deeply fuse and learn organelle class-specific features. [42] and SAM-Path [38] are proposed to embed other backbones de￾signed for the same task into SAM to… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Overview of the class prompt encoder. We present a framework without manual prompts that constructs a cosine similarity matrix to activate class features and generate dense embeddings with a residual connection structure. 3.3 Cosine Similarity based Class Prompt Encode…
Figure 5
Figure 5. Figure 5: Visual comparison of prediction masks on high- and low-glucose samples. These methods validated one islet cell in each dataset by overlaying the original images and segmentation masks. Notably, the yellow dashed ellipse is used to emphasize regions with significant rec…
Figure 6
Figure 6. Figure 6: Visualization of positive class similarity maps for three categories: (a) nucleus; (b) mitochondria; (c) granules. We evaluate the cosine similarity between image embeddings and trained class prototypes on a pixel-by-pixel basis and overlay the similarity matrix over t…
Figure 7
Figure 7. Figure 7: t-SNE discrete visualization embedding distribution for the high-glucose dataset: (a) Original distributions; (b) Aligned distributions. In addition, ScSAM can distinctly differentiate between the three organelles and excels in edge detection of mitochondria and granul…
Figure 8
Figure 8. Figure 8: Dice score over 30 epochs for comparison methods trained on high- and low-glucose BetaSeg datasets. We compare robustness and convergence speed by visualizing the evolution of the mean Dice score. mon feature space. 4.5 Efficiency Comparisons To comprehensively evaluat…
Figure 1
Figure 1. Figure 1: Visual comparison of prediction masks on high-glucose samples. These methods validated one islet cell in each dataset by overlaying the original images and segmentation masks. Notably, the yellow dashed ellipse is used to emphasize regions with significant recognition …
Figure 2
Figure 2. Figure 2: Visual comparison of prediction masks on low-glucose samples [PITH_FULL_IMAGE:figures/full_fig_p011_2.png]
Figure 3
Figure 3. Figure 3: Visualization of positive class similarity maps for three categories: (a) nucleus; (b) mitochondria; (c) granules. We evaluate the cosine similarity between image embeddings and trained category prototypes on a pixel-by-pixel basis and overlay the similarity matrix ove…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 38 canonical work pages

  1. [1]

    Abdur Rahman, I

    L. Abdur Rahman, I. Papathanail, L. Brigato, and S. Mougiakakou. A SAM based tool for semi-automatic food annotation. In Proceedings of the European Conference on Artificial Intelligence , pages 4475–4478. IOS Press, 2024

    work page 2024

  2. [2]

    Bissoto et al

    A. Bissoto et al. Debiasing skin lesion datasets and models? not so fast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 740–741, 2020

    work page 2020

  3. [3]

    Boykov et al

    Y . Boykov et al. Graph cuts and efficient N-D image segmentation. International Journal of Computer Vision, 70(2):109–131, 2006

    work page 2006

  4. [4]

    T. F. Chan and L. A. Vese. Active contours without edges. IEEE Trans- actions on Image Processing, 10(2):266–277, 2001

    work page 2001

  5. [5]

    J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, and Y . Zhou. TransUnet: Transformers make strong encoders for medi- cal image segmentation. arXiv preprint arXiv:2102.04306, 2021

    arXiv 2021

  6. [6]

    T. Chen, L. Zhu, C. Deng, R. Cao, Y . Wang, S. Zhang, Z. Li, L. Sun, Y . Zang, and P. Mao. SAM-Adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 3367–3375, 2023

    work page 2023

  7. [7]

    Cheng, Q

    Z. Cheng, Q. Wei, H. Zhu, Y . Wang, L. Qu, W. Shao, and Y . Zhou. Unleashing the potential of SAM for medical adaptation via hierarchical decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3511–3522, 2024

    work page 2024

  8. [8]

    J. Fan, D. Liu, H. Chang, H. Huang, M. Chen, and W. Cai. Taxonomy adaptive cross-domain adaptation in medical imaging via optimization trajectory distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21174–21184, 2023

    work page 2023

Show all 42 references
  1. [9]

    J. Fan, D. Liu, H. Chang, H. Huang, M. Chen, and W. Cai. Seeing unseen: Discover novel biomedical concepts via geometry-constrained probabilistic modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11524–11534, 2024

    2024

  2. [10]

    J. Fan, D. Liu, C. Li, H. Chang, H. Huang, F. Braet, M. Chen, and W. Cai. Revisiting adaptive cellular recognition under domain shifts: A contextual correspondence view. In European Conference on Com- puter Vision, pages 275–292. Springer, 2024

    2024

  3. [11]

    Han et al

    H. Han et al. Self-supervised voxel-level representation rediscovers sub- cellular structures in volume electron microscopy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1874–1883, 2022

    2022

  4. [12]

    Z. He, M. Unberath, J. Ke, and Y . Shen. TransNuSeg: A lightweight multi-task transformer for nuclei segmentation. InMedical Image Com- puting and Computer Assisted Intervention – MICCAI 2023, pages 206–

    2023

  5. [13]

    Heinrich, D

    L. Heinrich, D. Bennett, D. Ackerman, W. Park, J. Bogovic, N. Eck- stein, A. Petruncio, J. Clements, S. Pang, C. S. Xu, et al. Whole-cell organelle segmentation in volume electron microscopy. Nature, 599 (7883):141–146, 2021

    2021

  6. [14]

    Isensee, P

    F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203–211, 2021

    2021

  7. [15]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, et al. Segment any- thing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4026, 2023

    2023

  8. [16]

    Li et al

    X. Li et al. LoGoNet: Towards accurate 3D object detection with local- to-global cross-modal fusion. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 17524–17534, 2023

    2023

  9. [17]

    J. Lin, L. Liu, D. Lu, and K. Jia. SAM-6D: Segment anything model meets zero-shot 6D object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 27906–27916, 2024

    2024

  10. [18]

    Liu et al

    D. Liu et al. PDAM: A panoptic-level feature alignment framework for unsupervised domain adaptive instance segmentation in microscopy images. IEEE Transactions on Medical Imaging, 40(1):154–165, 2020

    2020

  11. [19]

    N. Luo, R. Sun, Y . Pan, T. Zhang, and F. Wu. Electron microscopy images as set of fragments for mitochondrial segmentation. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 3981–3989, 2024

    2024

  12. [20]

    J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024

    2024

  13. [21]

    Müller, D

    A. Müller, D. Schmidt, C. S. Xu, S. Pang, J. V . D’Costa, S. Kretschmar, C. Münster, T. Kurth, F. Jug, M. Weigert, et al. 3D FIB-SEM recon- struction of microtubule–organelle interaction in whole primary mouse β cells. Journal of Cell Biology, 220(2), 2021

    2021

  14. [22]

    Oktay et al

    O. Oktay et al. Attention U-net: Learning where to look for the pan- creas. arXiv preprint arXiv:1804.03999, 2018

    2018 arXiv

  15. [23]

    N. Otsu. A threshold selection method from gray-level histograms. Au- tomatica, 11(285-296):23–27, 1975

    1975

  16. [24]

    Z. Peng, Z. Xu, Z. Zeng, X. Yang, and W. Shen. SAM-PARSER: Fine- tuning SAM efficiently by parameter space reconstruction. In Proceed- ings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4515–4523, 2024

    2024

  17. [25]

    Ravi et al

    N. Ravi et al. SAM 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  18. [26]

    A. Rizk, G. Paul, P. Incardona, M. Bugarski, M. Mansouri, A. Niemann, U. Ziegler, P. Berger, and I. F. Sbalzarini. Segmentation and quantifica- tion of subcellular structures in fluorescence microscopy images using squassh. Nature Protocols, 9(3):586–596, 2014

    2014

  19. [27]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional net- works for biomedical image segmentation. In Medical Image Comput- ing and Computer-Assisted Intervention – MICCAI 2015 , pages 234–

    2015

  20. [28]

    A. A. Sekh, I. S. Opstad, G. Godtliebsen, Å. B. Birgisdottir, B. S. Ahluwalia, K. Agarwal, and D. K. Prasad. Physics-based machine learning for subcellular segmentation in living cells. Nature Machine Intelligence, 3(12):1071–1080, 2021

    2021

  21. [29]

    Y . Song, J. Fan, H. Huang, M. Chen, and W. Cai. Cell as Point: One-stage framework for efficient cell tracking. arXiv preprint arXiv:2411.14833, 2024

    2024 arXiv

  22. [30]

    Van der Maaten and G

    L. Van der Maaten and G. Hinton. Visualizing data using t-SNE. Jour- nal of Machine Learning Research, 9(11), 2008

    2008

  23. [31]

    Wang, W.-L

    Y . Wang, W.-L. Chao, D. Garg, B. Hariharan, M. Campbell, and K. Q. Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pa...

    2019

  24. [32]

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon. CBAM: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision, pages 3–19, 2018

    2018

  25. [33]

    K. Xie, R. Guo, C. Cong, M. Pagnucco, and Y . Song. Domain gener- alised cell nuclei segmentation in histopathology images using domain- aware curriculum learning and colour-perceived meta learning. In Pro- ceedings of the European Conference on Artificial Intelligence , pages...

    2024

  26. [34]

    R. Xie, K. Pang, G. D. Bader, and B. Wang. MAESTER: Masked au- toencoder guided segmentation at pixel resolution for accurate, self- supervised subcellular structure recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3292–...

    2023

  27. [35]

    Yamagiwa et al

    H. Yamagiwa et al. Zero-shot edge detection with SCESAME: Spectral clustering-based ensemble for segment anything model estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 541–551, 2024

    2024

  28. [36]

    S. Yuan, L. Luo, Z. Hui, C. Pu, X. Xiang, R. Ranjan, and D. Deman- dolx. UnSAMFlow: Unsupervised optical flow fuided by segment any- thing model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19027–19037, 2024

    2024

  29. [37]

    W. Yue, J. Zhang, K. Hu, Y . Xia, J. Luo, and Z. Wang. SurgicalSAM: Efficient class promptable surgical instrument segmentation. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 6890–6898, 2024

    2024

  30. [38]

    Zhang et al

    J. Zhang et al. SAM-Path: A segment anything model for semantic seg- mentation in digital pathology. In Medical Image Computing and Com- puter Assisted Intervention – MICCAI 2023 Workshops, pages 161–170. Springer, 2023

    2023

  31. [39]

    Zhang, T

    P. Zhang, T. Yan, Y . Liu, and H. Lu. Fantastic animals and where to find them: Segment any marine animal with dual SAM. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 2578–2587, 2024

    2024

  32. [40]

    Zhang, X

    Q. Zhang, X. Liu, W. Li, H. Chen, J. Liu, J. Hu, Z. Xiong, C. Yuan, and Y . Wang. Distilling semantic priors from SAM to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25409–25419, 2024

    2024

  33. [41]

    H.-Y . Zhou, J. Guo, Y . Zhang, L. Yu, L. Wang, and Y . Yu. nnFormer: Interleaved transformer for volumetric segmentation. arXiv preprint arXiv:2109.03201, 2021

    2021 arXiv

  34. [42]

    S. Zhou, R. He, W. Tan, and B. Yan. SAMFlow: Eliminating any frag- mentation in optical flow with segment anything model. In Proceed- ings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7695–7703, 2024. Supplementary Material 1 Implementation Details 1.1 P...

    2024

Pith tools

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