List of Papers Browse by Subject Areas Author List
Abstract
Cell segmentation is critical for computational pathology and biomedical discovery. While recent Vision Foundation Models (VFMs) have demonstrated remarkable universal feature representations, unlocking their full potential for cellular imaging is currently bottlenecked by resource-intensive adaptation paradigms. Existing methods typically rely on fine-tuning heavy visual encoders, leading to extensive computational overhead and a dependency on large-scale annotations. To address this, we propose the EffiCell-Seg framework for highly efficient cell segmentation without re-training the visual encoder. Our core insight is that pretrained VFMs intrinsically encode complementary structural priors: global saliency for localizing potential cells, and local morphological patterns for delineating cellular structures. To harness these priors, we devise a Cell Structure Prompt Encoder (CSP-Encoder) that synthesizes semantic-aware saliency and principal morphological features from frozen VFM representations into explicit structural prior maps. Moreover, we propose a Synergistic Mask Decoder (SM-Decoder) that enforces contextual consistency by jointly predicting geometric distance fields and semantic maps via mutual cross-guidance. Extensive experiments demonstrate that EffiCell-Seg outperforms state-of-the-art methods across diverse cell imaging modalities while requiring only $\sim$5M trainable parameters, over 130$\times$ fewer than fully fine-tuned VFM counterparts. The code is available at https://anonymous.4open.science/r/EffiCell-Seg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0247_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xq141839/EffiCell-Seg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XuQin_Rethinking_MICCAI2026,
author = { Xu, Qing AND He, Xiangjian AND Duan, Wenting AND Luo, Jiebo AND Chen, Zhen},
title = { { Rethinking the Adaptation of Vision Foundation Models for Efficient Cell Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
- This paper proposes a parameter-efficient cell segmentation framework that leverages the structural knowledge of Vision Foundation Models (VFMs) without requiring backbone fine-tuning.
- This paper introduces a CSP-Encoder, which utilizes the saliency map between spatial tokens and the CLS token, alongside the PCA principal components of the spatial tokens, to effectively bridge pretrained knowledge with cellular structures.
- SM-Decoder is proposed to jointly predict geometric distance fields and semantic maps via interactions between structure-aware prompt embeddings and spatial tokens.
- The proposed method, EffiCell-Seg, achieves strong performance while using substantially fewer trainable parameters than prior approaches.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.This paper achieves highly competitive performance while requiring significantly fewer trainable parameters compared to existing baseline approaches.
- The ablation results are informative and support the claim that both the CSP-Encoder and the SM-Decoder play meaningful roles in the observed performance gains. 3.Since the proposed framework operates on frozen VFM features rather than relying on backbone-specific fine-tuning, it appears potentially extensible to other VFM architectures.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The paper emphasizes computational efficiency, yet the evidence is limited to trainable parameter counts. Reporting latency, GPU memory, or throughput under the same hardware would substantially strengthen this claim. 2.The paper states that PCA is applied to the spatial tokens and that the top-K components are retained, but it does not report the actual value of K and analyze the sensitivity to this hyperparameter. Since PCA-based morphology is a central part of the method, an ablation over K would improve both clarity and reproducibility. 3.It remains unclear how much of the reported superiority comes from the proposed design versus the use of a strong frozen DINOv3 backbone. 4.The paper identifies reliance on large-scale annotated data as a major limitation of prior work; however, the proposed method also appears to require considerable annotated supervision, given that it is trained with ground-truth.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposes an interesting parameter-efficient framework for cell segmentation by leveraging the structural knowledge of frozen Vision Foundation Models (VFMs). However, the claim regarding computational efficiency is insufficiently supported, as it relies solely on trainable parameter counts without providing practical metrics such as latency or GPU memory usage. Furthermore, while the authors identify the dependency on large-scale annotations as a major limitation of prior work, the proposed method itself still requires dense pixel-level ground-truth masks for training.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed my concerns in the rebuttal. I am changing my recommendation to Accept.
Review #2
- Please describe the contribution of the paper
This paper proposes EffiCell-Seg, an efficient cell segmentation framework that leverages frozen visual foundation models (VFM) without fine-tuning the visual encoder. The method is motivated by the observation that pretrained VFM contains useful structural priors, including global saliency cues for cell localization and local morphological patterns for boundary delineation. To exploit these cues, the authors propose Cell Structure Prompt Encoder (CSP-Encoder) to extract the structural prior maps from frozen VFM features and a Synergistic Mask Decoder (SM-Decoder) to jointly predict semantic maps and geometric distance fields through cross-guidance. Experimental results on the CellSeg and DSB datasets show that the method achieves strong performance in both binary and instance segmentation with smaller parameter count.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1) This method is motivated by a sensible and potentially useful observation. The authors identify that despite the domain gap, pretrained VFM models may remain useful for cell segmentation since cell instances also exhibit structured boundaries and locally coherent regions. The qualitative evidence in Figure 1 strengthens this motivation. In addition, the use of a frozen image encoder is practically appealing, as it substantially reduces computational cost during training.
2) The reported results show promising results against prior methods on both binary and instance segmentation, while maintaining a much smaller parameter count.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1) Some implementation details could be described more clearly, particularly the neck layer and its role.
2) The construction of geometric and semantic queries is not fully explained, and additional detail would improve the clarity and reproducibility of the method.
3) Since the proposed framework relies heavily on pretrained VFM features, an additional comparison or ablation across different types of pretrained VFMs, such as SAM, would help clarify how sensitive the method is to the choice of foundation model and would provide a better understanding of how much the performance depends on the selected pretrained backbone.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is overall well-motivated. My concerns are relatively minor and mainly related to the clarity of several architectural details and supporting ablations.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors addressed my concerns, particularly those regarding the explanation of several components and missing details that i had raised, as well as other reviewers’. Although one of my concern is left for future work, it overlaps with a point raised by R3 that was adequately addresses in the rebuttal. Overall, i maintain my recommendation for acceptance.
Review #3
- Please describe the contribution of the paper
EffiCell-Seg adapts a frozen VFM for cell segmentation by extracting structural priors from DINOv3 features using CLS-token cosine similarity and PCA of spatial tokens, fusing them into structural prior maps via a lightweight CSP-Encoder. A Synergistic Mask Decoder then jointly predicts geometric distance fields and semantic maps through bidirectional cross-attention, with watershed post-processing for instance separation. Only ~5M parameters are trained.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The core idea is clean: frozen VFM features already encode global saliency and local morphology, and re-aligning these is sufficient without any encoder fine-tuning. The 130x parameter reduction over CellSAM while still outperforming it is compelling. The ablation honestly isolates each component, and the gains are consistent across both datasets and both task types.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The method is only tested with DINOv3 as the backbone and never ablates whether the claimed structural priors generalize to other VFMs like SAM or CLIP. The two datasets are small and familiar; results on PanNuke or CoNIC would strengthen generalization claims. Performance margins over the second-best method are modest, and no variance across runs is considered. Minor: The DINOv3 citation points to an August 2025 arXiv preprint, which raises questions about which model version was actually used!
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The frozen-encoder angle is interesting, and the efficiency part is strong, but the single-backbone dependency without ablation and limited dataset diversity leaves many open questions.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
R1Q1: Computational efficiency. A: On a single NVIDIA H20 GPU at 512×512 resolution, EffiCell-Seg achieves 54 ms latency, 18.5 FPS throughput, and 3.5 GB GPU memory. In contrast, all SAM-based methods in Tab 1-2 using heavier SAM-ViT-H require 115-178 ms latency, 5.6-8.7 FPS throughput, and 6.4-8.7 GB memory. This confirms that EffiCell-Seg is not only parameter-efficient but also achieves superior inference efficiency over existing methods.
R1Q2: PCA hyperparameter K. A: Ablation with K=1/2/3 yields Dice 87.70/87.71/87.76 on CellSeg and 93.14/93.17/93.21 on DSB, with corresponding latency of 54/57/61 ms. Performance differences are within 0.1% while latency grows linearly with K. We therefore adopt K=1 as the best trade-off between performance and computational cost.
R1Q3: Source of superiority. A: As illustrated in Tab 3, directly decoding frozen DINOv3 features yields only 83.52/89.74 Dice on CellSeg/DSB. Adding our CSP-Encoder and SM-Decoder brings 4.18/3.40 Dice gains to 87.70/93.14, outperforming all baselines. This confirms that the reported superiority primarily stems from our proposed design.
R1Q4: Annotation requirement. A: EffiCell-Seg does not require considerable annotated supervision. By freezing the VFM encoder and only optimizing the lightweight CSP-Encoder and SM-Decoder (~5M parameters), our method achieves superior performance with only a few hundred training samples. In contrast, methods like CellSAM [16] fully fine-tune heavy encoders on more than 10K annotated cell images. Our design greatly reduces the labeled samples required for adapting VFMs to cell segmentation.
R2Q1: Neck layer. A: The neck layer consists of a 1×1 convolution, LayerNorm, 3×3 convolution, and LayerNorm, which projects frozen VFM spatial tokens into a unified feature space aligned with the prompt embedding dimension for subsequent fusion in the SM-Decoder (Eq. 4).
R2Q2: Geometric and semantic queries. A: As shown in Fig. 2 and Eq. 4-5, q_g and q_s are randomly-initialized learnable tensors optimized end-to-end: q_g captures spatial organization for distance field prediction, while q_s encodes category-aware features for semantic mapping. They are first grounded via two-way attention with VFM features and the prompt embedding (Eq. 4), then mutually refined through bidirectional cross-attention (Eq. 5).
R2Q3: Comparison across VFMs. A: We agree that exploring our framework with other VFMs would further strengthen the generality, and will explore this in future work.
R3Q1: Empirical VFM comparison. A: Replacing DINOv3 with SAM and CLIP as the frozen backbone, EffiCell-Seg achieves 87.02/92.48 and 85.93/91.37 Dice on CellSeg/DSB, respectively, compared to 87.70/93.14 with DINOv3.DINOv3 provides the best structural priors for cell segmentation, while the framework remains effective across different VFMs.
R3Q2: Dataset diversity. A: Our CellSeg and DSB datasets cover fluorescence, brightfield, DIC, and H&E modalities with multiple cell types. In contrast, the suggested PanNuke and CoNIC are both restricted to H&E only. Our datasets exhibit greater modality diversity.
R3Q3: Performance margins and variance. A: Although the PQ gains over the second-best SAC are 1.27-1.38 on the two datasets, paired t-tests across test samples confirm strong statistical significance with p-values < 0.001 on both CellSeg and DSB. Moreover, the consistency across all four metrics (Dice, mIoU, PQ, F1) demonstrates that our improvements are robust rather than attributable to run-to-run variance.
R3Q4: DINOv3 version. A: We confirm that we use the August 2025 arXiv preprint version of DINOv3 [22] (the first official release), which is also the version publicly available on the official repository.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper received mixed feedback, and the reviewers raised concerns that I believe can be (partially) resolved. In particular, the two rejection scores raised concerns regarding missing implementation details and insufficient ablation justification, which are well-suited for clarification in a rebuttal. Also, the reviewers appreciate the novelty and strong performances. Thus, I believe this submission deserves an opportunity to rebut.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers appreciate the novelty and strong performances, and the initial concerns have been fully addressed during the rebuttal. I recommend accepting this submission.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper presents a novel and practical approach to efficient cell segmentation by re-aligning structural priors from a frozen VFM using the proposed CSP-Encoder and SM-Decoder, achieving strong semantic and instance segmentation results with only ~5M trainable parameters. Although initial concerns were raised about implementation details, efficiency evidence, backbone dependence, and ablations, the rebuttal satisfactorily clarified most key points, and two reviewers updated or maintained accept recommendations; remaining limitations do not outweigh the paper’s novelty, efficiency, and empirical strength.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All reviewers recommended acceptance, and I see no compelling reason to recommend otherwise.