acceptodds
Under review as a conference paper at ICLR 2027

Type-Aware Representation Learning with Dual-Teacher Consistency for Label-Efficient Nuclei Segmentation

Abstract

Nuclei instance segmentation in histopathology requires precise separation of touching nuclei and accurate cell-type classification. However, dense annotations are costly, and compact networks must preserve fine structural details while learning discriminative semantic representations. To address these challenges, we propose SPiN, a compact multitask architecture, together with dual-teacher Mean Teacher training (DTMT) for image-level semi-supervised nuclei instance segmentation. SPiN integrates a FastViT encoder, a parallel local enhancement pathway, and a shared decoder, with auxiliary type supervision to strengthen semantic representations. DTMT employs two exponential moving average teachers with distinct update rates to capture complementary temporal information. Their type predictions are combined through confidence-weighted fusion and filtered to provide soft supervision for unlabeled images, while geometric displacement regression remains supervised by labeled instances. Only the student network is required at inference. On PanNuke, the framework uses 12.50 million parameters and achieves bPQ/mPQ scores of 0.558/0.396 with 10% labeled training images, compared with 0.488/0.304 for supervised-only training and 0.553/0.390 for an adapted UniMatch baseline. These results demonstrate that integrating compact multitask modeling with type-focused temporal consistency improves annotation efficiency while maintaining a compact inference model, providing an effective approach to nuclei instance segmentation under limited supervision.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
semi-supervised learning, nuclei instance segmentation, teacher-student learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_YFNoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPiN with DTMT, which leverages type-aware representations and confidence-weighted dual EMA teachers to improve label-efficient nuclei instance segmentation.

Critical strengths

The architecture is compact and practically relevant, achieving competitive accuracy with low inference cost.

Critical weaknesses

1.The methodological novelty is limited. Multiple/temporally diverse teacher models have already been studied, including Switching Temporary Teachers and Progressive Mean Teachers, but the paper does not adequately discuss or compare against these closely related methods. 2.The contributions of the second teacher and different EMA rates are not adequately ablated.

Questions for authors

1.Can the authors separately ablate the second teacher, unequal EMA rates? 2.Could the authors standardize the border thickness and connector rendering in Figure 3, particularly around the “TP Branch Consistency Loss” box?

Reviewer_GqqKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SPiN, a multi-task segmentation framework, and DTMT, a semi-supervised learning strategy, to reduce labeling/annotation costs for cell panoptic segmentation (i.e., instance segmentation with instance-level cell-type classification). SPiN integrates a FastViT encoder with a NuLite-based decoder and employs Hover-Net-style output formats, such as foreground, HV, and type prediction. In the proposed SPiN, LAEU also enhances local features, while TASP adds auxiliary type supervision to the lowest-resolution features like common segmentation protocols (e.g., UPerNet).

Specifically, DTMT is based on EMA-based semi-supervised learning by combining type probabilities from two teachers with different EMA rates using confidence weights and applying consistency loss to unlabeled pixels above a confidence threshold. SPiN is evaluated for annotation budgets and panoptic quality on PanNuke, Lizard, and CoNSeP.

Critical strengths
  • Reducing annotation costs for instance segmentation and instance-level cell-type classification is practical.

  • Similar to existing semi-supervised methods, inference uses only the student for a efficient deployment. The 12.50M-parameter segmentation model uses an existing/predefined FastViT encoder and provides empirical accuracy-efficiency results.

Critical weaknesses
  • [W1] Weak Novelty / Missing Relevant Semi-Supervised Segmentation Results. Applying well-known EMA, confidence-based pseudo-labeling, and auxiliary supervision to cell panoptic segmentation does not by itself establish a new contribution. Although Table 2 includes adapted CPS/UniMatch baselines, comparisons and distinctions from the cited (see L104) BASS and UniMatchv2 remain unclear. Component-wise differences from related dual-teacher fusion, reliability-weighted soft-target consistency, and mask-quality filtering methods [A, B, C] are also not compared. Please compare conceptual advances from differences in tasks and supervision, and validate teacher updates, and instance/type supervision through controlled comparisons. In the current paper, the main contribution appears limited to application novelty and empirical results.

  • [W2] Unclear Reproducibility / Comparison Protocols. Tables 1/2 must isolate semi-supervised training strategies, while Tables 3/4 must evaluate fully-supervised architectures. Details on shared encoders/decoders, pretrained backbones, dataset splits, annotation budgets, and FixMatch/CPS/UniMatch adaptation are insufficient, and no code or appendix is provided to assess reproducibility. Also, stability across seeds is important at high unlabeled ratios (e.g., >80%). Please clarify whether the fully-supervised Hover-Net/CellViT/NuLite-H results use training configurations (e.g., dataset splits). As shown in Table 4, the source of the parameter gap between NuLite-H and SPiN is unclear. Since FastViT is simply adopted, architecture choices and semi-supervised contributions must be ablated separately. Reporting comparable performance while presenting existing components as contributions further weakens the novelty claim.

  • [W3] Fixed Hyperparameters and Training Choices. Sensitivity analysis is needed for the consistency weight of 0.02 and 2000-step in Eq. 4, EMA rates of 0.99/0.995, and confidence threshold of 0.7. No experiment establishes whether this combination remains effective across annotation budgets. At minimum, compare a single teacher, two teachers with identical or different EMA rates, and uniform versus confidence-weighted fusion [A, B]. Please also justify excluding NP/HV from unlabeled consistency. The current ablation shows the experimental justification incomplete and unclear individual contributions/gains.

  • [W4] Encoder and Decoder Choices. The choice of FastViT and NuLite's decoder needs accuracy-efficiency comparisons against similarly sized (or smaller) alternatives. Mask decoders and encoder-only methods, such as EoMT [D] should be considered in terms of training and inference costs. Retaining Hover-Net-style HV prediction and watershed also requires positioning against recent alternatives (e.g., [E]).

  • [W5] Task Importance. Given the focus on annotation efficiency, please compare with training-free cell instance segmentation approaches such as CoP [F] and SPROUT [G] to justify the need for task-specific training/supervision. Specifically, CoP requires a click per cell type at inference time but performs training-free cell panoptic segmentation. SPROUT does not directly provide cell-type classification. However, that joint cell panoptic segmentation offers any advantages over a two-stage approach (cell instance segmentation -> cell-type classification) is unclear. This result will clarify and enhance when the proposed semi-supervised training offers a practical advantage beyond empirical comparisons.

  • [W6] Readability. Please improve Figure 1's spacing and plot labels, and include key baselines such as CellViT in the qualitative comparison. Also, I suggest merging figures 2/3 into an overview showing labeled/unlabeled real training and inference samples, followed by separate descriptions of individual modules.

References:

[A] Perturbed and Strict Mean Teachers for Semi-Supervised Semantic Segmentation, CVPR 2022.

[B] Dual-teacher fusion with augmented branch for semi-supervised object detection, Pattern Recognition 2026.

[C] MUST: Mean-teacher inspired and Uncertainty-aware Semi-supervised Transformer for cell instance segmentation, JBHI 2026.

[D] Your ViT is Secretly an Image Segmentation Model, CVPR 2025.

[E] The Four Color Theorem for Cell Instance Segmentation, ICML 2025.

[F] One Click per Cell Type Suffices: Training-free Group Interaction for Cell Instance Segmentation, ACCV 2026.

[G] Supervise Less, See More: Training-free Nuclear Instance Segmentation with Prototype-Guided Prompting, ICML 2026.

Questions for authors

Please see the Weaknesses for details.

Q1. [W1 & W2 & W4] Were all baselines in Tables 1/2 evaluated with the same architecture, pretraining, dataset splits, annotation budgets, and training conditions as SPiN for a fair comparison? What explains the parameter difference NuLite-H and SPiN (e.g., only encoder)?

Q2. [W1 & W3] What are DTMT's conceptual distinctions from BASS, UniMatchv2, and related works [A, B, C]? Under matched training conditions, what controlled evidence isolates the benefits of dual EMA teachers and confidence-weighted fusion over strong single-teacher and uniform-fusion baselines?

Q3. [W5] What evidence supports the practical advantage of jointly learning instance segmentation and cell-type prediction over a two-stage pipeline or training-free alternatives [F, G]?

Reviewer_oXiDOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes to combine FastViT-based nuclei segmentation architecture with lcoal feature enhancement, auxiliary type supervision, and two EMA teachers. The proposed method is evaluated in instance quality and nuclear types across annotation budgets on two public datasets.

Critical strengths
  • The paper addresses a relevant problem in nuclei segmentation.
  • The writing is clear and the overall structure of the paper is well organized.
Critical weaknesses
  • The main contribution of the paper is not clearly stated. The idea of combining local feature enhancement, type supervision, and semantic supervision has been explored in existing works like HoverNet and NuLite. Using two EMA teachers has also been explored in Dual Teacher (Na et al., 2023) and HD-Teacher (Zhu, 2023). The authors should clarify what is novel in their approach compared to these existing methods.
  • The effectiveness of the DTMT method is not clearly demonstrated. The authors claim that using different EMA factors for the two teachers can capture temporal information. However, the two teachers are trained on the same model architecture and close EMA factors, which may not provide significant diversity in the learned representations. The authors should provide both theoretical and empirical evidence to support their claim that the two teachers can capture different temporal information, and that this leads to complementary benefits to the student model. Currently, the paper only provides empirical results showing that DTMT outperforms a single teacher, but it does not provide any analysis or ablation studies to demonstrate the complementary benefits of the two teachers.
  • Semi-supervised experiments need repeated sampling and uncertainty analysis. The paper does not specify how labeled images are sampled or report variation across different runs. This is important to semi-supervised settings, especially for the rare classes like dead-cells.
  • The improvement of the proposed method over the baseline is marginal. In Table 2, with 10% labeled data, the proposed method only improves the UniMatch baseline by 0.005 bPQ and 0.002 F1, which is not significant. The authors should provide a more detailed analysis of the results, including statistical significance tests, to demonstrate the effectiveness of their method.
  • Efficiency analysis is unconvincing. In Table 4, the proposed method is compared with NuLite-H, and shows significant improvement in efficiency. However, the original NuLite paper provides a more lightweight version, NuLite-T, which shows better efficiency than the proposed method in terms of parameters.

Ref:

  1. Na, et al. Switching Temporary Teachers for Semi-Supervised Semantic Segmentation. NeurIPS 2023.
  2. Zhu, et al. Hybrid Dual Mean-Teacher Network With Double-Uncertainty Guidance for Semi-Supervised Segmentation of MRI Scans. CMIG, 2024.
Questions for authors
  • The effectiveness of the dual-teacher design can be further clarified. Compared with a single teacher in different EMA factors, the necessity of using two teachers should be justified. Does introducing more teachers lead to better performance? If so, how many teachers are optimal? The fusion strategy in Eq. (2) should also be explained in more detail. Does the confidence-based fusion strategy provide better performance than a simple averaging of the two teachers' outputs? How to address the discrepancy between the two teachers' outputs when they disagree on certain predictions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.