acceptodds
Under review as a conference paper at ICLR 2027

LiNA: Rethinking Shape Capacity for Efficient Nuclei Analysis

Abstract

How much shape-decoding capacity does joint nuclei analysis actually need? High-performing systems devote substantial computation to predicting instance geometry, yet it is unclear whether this cost is justified by the complexity of nuclear shape. If nuclear geometry needs only a few degrees of freedom, dense shape decoding may spend much of that computation on capacity with little task-level benefit. We probe this by varying Fourier order N , which controls the boundary’s degrees of freedom. Across four histopathology datasets, a single Fourier harmonic reconstructs ground-truth contours at 85–89% IoU and order N =15 reaches 98.5–99.1%, with an independent boundary-distance measure showing the same low-order saturation. These results raise the possibility that current pipelines devote more capacity to nuclear geometry than the task requires. To test and exploit this observation, we instantiate LiNA (Lightweight Nuclei Analysis), which separates localization and typing from geometry prediction, represents each nucleus with a compact Fourier contour, and strengthens typing through prototype distillation. We then replace its Fourier pathway with a matched dense-mask decoder while keeping the remaining system fixed. Across four backbone scales, the dense-mask decoder improves none of the evaluated task metrics, while computation grows with nucleus density and reaches 6.4× that of the matched Fourier variant on the densest PanNuke tile for ConvNeXt-Small. This compact geometry pathway is nevertheless sufficient for strong joint analysis. At 57M parameters, roughly 8% of 700M-scale SAM-based pipelines, LiNA achieves the best performance on PanNuke and Lizard and the strongest zero-shot results on MoNuSeg and CPM17. Together, these results reveal a shape-capacity mismatch in joint nuclei analysis. LiNA exploits this mismatch to reduce geometry capacity without sacrificing recognition and, combined with its efficient overall design, reaches a new Pareto-efficient operating point among the compared models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
efficient deep learning, instance segmentation, computational pathology, nuclei analysis, Fourier contours, knowledge distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JJjpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how much shape-decoding capacity is necessary for joint nuclei detection, classification, and instance segmentation. Through Fourier contour fitting on four histopathological datasets, the author found that lower-order representations can achieve high shape reconstruction accuracy, and proposed LiNA: separating point localization, instance classification, and compact Fourier contour prediction, and combining curvature-aware boundary loss and prototype distillation to improve segmentation and classification. Validation includes Fourier-order sweeps, comparisons against a per-instance dense-mask decoder across four backbone scales, joint-analysis benchmarks on PanNuke and Lizard, and transfer without fine-tuning to MoNuSeg and CPM17. LiNA achieves a favorable accuracy–inference-efficiency trade-off among the evaluated systems.

Critical strengths
  1. This paper introduces a clear notion of “shape capacity” for nuclear contours, using Fourier order for representational complexity.
  2. This paper designs LiNA to decouple detection/typing from geometry, enabling controlled capacity studies that isolate the effect of the shape pathway.
  3. This paper proposes a practical curvature-aware boundary loss to preserve sharp details under low-order contour representations.
  4. This paper presents a prototype distillation scheme that transfers semantics from a frozen teacher into a compact codebook and fuses it with student features, yielding sizable classification gains.
Critical weaknesses
  1. The evidence for a general shape-capacity mismatch remains limited. The main controlled comparison uses a Mask R-CNN-style per-instance dense decoder, whose computational cost inherently grows with nucleus count. This establishes LiNA’s advantage over the tested design, but does not establish that dense shape decoding is generally over-provisioned. Comparisons against shared dense decoders or substantially lighter mask heads would better support the paper’s central claim.

  2. The conceptual advance beyond existing compact-contour methods needs clearer substantiation. CPN and FourierNet already establish the effectiveness of Fourier-based instance representations. LiNA adds a useful joint-analysis architecture, boundary supervision, and prototype distillation, but the current evidence more clearly demonstrates an effective system than a substantially new principle. In particular, detection and classification are architecturally separated from geometry prediction, so their stability across Fourier orders offers limited additional support for the broader shape-capacity argument.

  3. The overall performance gains are not fully disentangled from pretrained representations and distillation. LiNA benefits from DINOv3 initialization and pathology-teacher supervision, while its comparisons against larger systems involve different architectures and training recipes. Table 6 usefully isolates CAB and distillation gains within LiNA, but does not establish how much of the advantage over competing systems comes from compact geometry rather than these other factors. Table 8 provides encouraging cross-dataset segmentation results, although the conclusions about transferable joint-analysis performance remain limited without corresponding detection and fine-grained classification evaluation.

Questions for authors
  1. A comparison against a shared dense decoder or a substantially lighter per-instance mask head is needed to substantiate the broader shape-capacity mismatch claim. Without such evidence, the conclusion should be restricted to the tested decoding pathways.

  2. The consistency of source training splits, checkpoint-selection procedures, and image resolutions across the models in Table 8 should be clarified. Evidence of cross-domain fine-grained classification or joint-analysis performance is also needed to establish whether LiNA’s advantage extends beyond binary instance segmentation.

  3. Comparisons with matched pretrained initialization and teacher supervision are needed to establish whether LiNA retains its advantage over larger systems. The contributions of compact geometry, backbone pretraining, and prototype distillation should be distinguished more clearly.

6 comments
AuthorsOpenReview ↗

Q1. Validating Shape-Capacity Mismatch with Shared and Lightweight Decoders (1/3)

Q1. Validating Shape-Capacity Mismatch with Shared and Lightweight Decoders (1/3)

As the reviewer pointed out, comparing LiNA only against the Mask R-CNN-style dense decoder in the original manuscript is insufficient to establish that shape-capacity mismatch is not specific to a particular decoder architecture. To address this concern, we introduced additional CondInst-style dynamic decoders, YOLACT-style shared pixel decoders, and Binary + Center-guided Watershed, along with experiments that increase the output mask resolution and RoI sampling resolution of the original dense decoder.

Our goal is not to claim that Fourier contours are the only effective lightweight representation, but rather to investigate whether the limited benefits of high shape-decoding capacity are consistently observed across different geometry decoder architectures.

1. Additional Decoder Architectures and Experimental Setup

We constructed the additional models by replacing LiNA's geometry prediction pathway while retaining the same backbone, point detection, classification, prototype distillation, and Hungarian matching.

The compared decoders are as follows:

  • Per-instance Dense Decoder: Based on a Mask R-CNN-style architecture, this decoder generates an independent dense mask from each nucleus's RoI features. Dense28 predicts a 28×28 mask from 14×14 RoIAlign features, while Dense56-A produces a 56×56 mask using the same RoIAlign resolution. Dense56-B further increases the RoIAlign resolution to 28×28 while predicting a 56×56 mask.
  • CondInst-style Dynamic Decoder: This decoder computes shared mask features at the image level and applies lightweight dynamic convolution filters generated from each nucleus's student RoI features to predict instance masks.
  • SharedPixel Decoder: This decoder generates K spatial prototype masks over the entire image and combines them using coefficients predicted from each nucleus's student RoI features. We evaluated K=8, 32, and 64 to examine the effects of shared representation capacity.
  • Binary + Center-guided Watershed: This approach predicts an image-level binary foreground map and separates instances using watershed with the detected nucleus centers as markers. It serves as a comparison method that performs instance separation without a separately learned per-instance shape output.

All additional decoders were evaluated on the same three train/validation/test splits of PanNuke. Decoders producing pixel-wise instance masks were trained using BCE and Dice losses, while Binary + Watershed was trained for binary foreground prediction and performed instance separation using settings selected on the validation set.

AuthorsOpenReview ↗

Q1. Validating Shape-Capacity Mismatch with Shared and Lightweight Decoders (2/3)

Q1. Validating Shape-Capacity Mismatch with Shared and Lightweight Decoders (2/3)

2. Performance Comparison Across Geometry Decoders

Table Q1-1. Performance comparison across geometry decoders (PanNuke, three-fold average).

ModelDet-F1Cls-F1avgPQ
LiNA (Ours)88.0871.9145.12
Dense2887.3870.4744.06
Dense56-A87.4670.3444.22
Dense56-B87.4270.3744.36
CondInst-style87.5070.8744.04
SharedPixel (K=8)87.4870.4543.30
SharedPixel (K=32)87.5070.5043.37
SharedPixel (K=64)87.4770.4343.51
Binary + Watershed87.7071.1942.41

For the per-instance dense decoders, increasing the mask resolution from 28×28 to 56×56 and further increasing the RoIAlign resolution improved avgPQ by only 0.30 percentage points, from 44.06 to 44.36. Similarly, increasing the number of spatial prototypes in SharedPixel eightfold, from K=8 to K=64, yielded only a 0.21 percentage-point improvement in avgPQ.

More importantly, CondInst-style and SharedPixel achieved competitive segmentation performance without high-capacity per-instance dense decoding. CondInst-style achieved an avgPQ of 44.04, only 0.32 percentage points below Dense56-B (44.36), while SharedPixel achieved avgPQ scores ranging from 43.30 to 43.51.

These results are also competitive with existing high-capacity nuclei analysis methods. For example, CellViT and PromptNucSeg achieved avgPQ scores of 43.1 and 44.3, respectively, as reported in our manuscript, but required approximately 943 and 570 GFLOPs. In contrast, CondInst-style and SharedPixel achieved comparable levels of accuracy using only approximately 58–63 GFLOPs, corresponding to 9–16× fewer total FLOPs. We note that these comparisons concern the efficiency of complete systems and do not attribute the performance differences solely to their geometry decoders.

These findings show that increasing decoding capacity provides limited performance improvements in both Dense and SharedPixel decoders, while other lightweight geometry decoders, such as CondInst-style, can also achieve competitive segmentation accuracy. This supports our observation that high shape-decoding capacity is not necessarily required for high accuracy and that this finding is not specific to a single decoder architecture.

Binary + Watershed achieved the lowest avgPQ among the compared methods, at 42.41. Unlike LiNA and the other lightweight decoders, which generate individual shape predictions for each nucleus, this approach separates nuclei using a global binary foreground mask and detected nucleus centers. These results support the effectiveness of reducing shape-decoding computation while retaining the ability to predict individual nucleus shapes.

AuthorsOpenReview ↗

Q1. Validating Shape-Capacity Mismatch with Shared and Lightweight Decoders (3/3)

Q1. Validating Shape-Capacity Mismatch with Shared and Lightweight Decoders (3/3)

3. Computational Efficiency of Geometry Decoding

To compare computational costs across decoders, we measured the total model FLOPs, geometry branch FLOPs, and geometry branch GPU execution time.

Table Q1-2. Computational costs of geometry decoders (multipliers relative to LiNA in parentheses). Total FLOPs are measured with THOP (FLOPs = 2 × MACs) on a single 256×256 input with all 256 candidate RoIs decoded, using the same counting as Table 5(b). Multipliers are computed from unrounded values.

ModelTotal FLOPs (G)Geometry FLOPs (G)Geometry GPU Time (ms)
LiNA (Ours)48.0 (1.0×)0.042 (1.0×)0.23 (1.0×)
Dense28395.6 (8.2×)347.63 (≈8,300×)10.61 (46.1×)
Dense56-A816.8 (17.0×)768.84 (≈18,400×)21.78 (94.7×)
Dense56-B1,421.8 (29.6×)1,373.89 (≈32,900×)39.42 (171.4×)
CondInst-style58.3 (1.2×)10.41 (≈250×)9.14 (39.7×)
SharedPixel (K=8)63.2 (1.3×)15.22 (≈365×)1.43 (6.2×)
SharedPixel (K=32)63.2 (1.3×)15.27 (≈366×)1.60 (7.0×)
SharedPixel (K=64)63.3 (1.3×)15.34 (≈368×)1.98 (8.6×)
Binary + Watershed58.3 (1.2×)10.37 (≈249×)0.70 (3.0×)

Dense56-B required approximately 3.6× more total FLOPs than Dense28, yet improved avgPQ by only 0.30 percentage points. In contrast, CondInst-style and SharedPixel shared most of their computation at the image level and achieved competitive segmentation accuracy with approximately 22–24× fewer total FLOPs than Dense56-B.

In particular, these shared decoders incur relatively little additional computation per instance, limiting the growth in total computational cost as the number of nuclei increases. This further reinforces our observation that high-capacity per-instance dense decoding is not essential for nuclei segmentation.

Meanwhile, Fourier contours offer an additional advantage over these lightweight decoders by substantially reducing the computational cost of the geometry prediction pathway itself.

LiNA's geometry branch requires only 0.042 GFLOPs, approximately 1/250 of CondInst-style and 1/365 of SharedPixel. Its geometry branch GPU execution time is also only 0.23 ms, approximately 39.7× faster than CondInst-style and 6.2–8.6× faster than SharedPixel.

This efficiency arises because Fourier contours directly predict each nucleus's shape using compact Fourier coefficients, without requiring separate spatial mask feature generation or dynamic convolutions. However, the differences in total model FLOPs relative to CondInst-style and SharedPixel are approximately 1.2–1.3×. We therefore distinguish the substantial computational savings in the geometry branch from the overall inference efficiency of the complete model.

4. Implications for Shape-Capacity Mismatch

Our central claim is not that Fourier contours are the only effective lightweight geometry representation, but that high shape-decoding capacity provides limited benefits for modeling nucleus geometry.

Beyond the low representational complexity of ground-truth contours and the Fourier-versus-dense comparison presented in the original manuscript, these additional experiments demonstrate that increasing the capacity of Dense and SharedPixel decoders yields only limited performance gains, while alternative lightweight decoders such as CondInst-style can also achieve competitive accuracy. In contrast, the lower performance of Binary + Watershed highlights the importance of effective instance separation rather than simply eliminating geometry computation.

Motivated by these observations, LiNA separates detection/classification from geometry prediction and achieves high segmentation accuracy with a highly compact Fourier contour representation and minimal geometry-decoding cost. These additional results therefore support that shape-capacity mismatch is not limited to the comparison with the original Mask R-CNN-style decoder.

Reviewer_s9hEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LiNA, a lightweight framework for nuclei instance segmentation that localizes nuclei through point detection and represents instance contours using Fourier coefficients. It incorporates a Curvature-Aware Boundary loss and prototype distillation to improve segmentation and classification performance. Through Fourier order analysis, dense-mask decoder comparisons, and cross-dataset experiments, the authors investigate whether low-order Fourier representations can maintain task performance while reducing computational costs.

Critical strengths

The paper presents a relatively novel perspective by questioning whether high-capacity dense shape decoding is necessary for nuclei instance segmentation. Its analysis of shape representation capacity provides an interesting alternative to conventional mask-based approaches and offers useful insights into the trade-off between segmentation performance and computational efficiency.

Critical weaknesses
  1. The mask head uses a fixed 28×28 resolution and a different training objective, so it is unclear whether the results reflect limited benefits of dense shape decoding or simply the specific baseline design.
  2. Most key comparisons lack results uncertainty estimates. Given the small performance differences, it is difficult to conclude that dense-mask decoding offers no real benefit.
  3. Reconstructing ground-truth contours shows that low-order Fourier representations can approximate nuclear shapes, but does not establish their reliability on difficult cases, such as crowded nuclei or irregular boundaries.
  4. The overall performance may also benefit from prototype distillation, CAB loss, and pretrained features. The current ablations do not clearly separate these effects, making it difficult to assess how much the Fourier representation itself contributes.
Questions for authors
  1. Would the conclusions hold with higher-resolution mask outputs or alternative mask losses under a more comparable experimental setup?
  2. Can the authors report more results, including standard deviations or confidence intervals?
  3. Can the authors evaluate contour prediction on challenging cases, such as crowded nuclei and irregular boundaries, rather than relying mainly on ground-truth contour reconstruction?
  4. Can the authors provide separate ablations for Fourier representation, prototype distillation, fusion, and CAB loss to clarify the contribution of each component to the overall performance?
7 comments
AuthorsOpenReview ↗

Q1. Higher-Resolution Masks and Alternative Mask Objectives

We sincerely thank the reviewer for the careful evaluation of our manuscript and the thoughtful, constructive feedback. We particularly appreciate the suggestions for strengthening our experimental analysis and examining alternative explanations for the observed results. These comments motivated us to conduct additional controlled experiments and helped us provide a more thorough and convincing evaluation of our central findings. We address each question in detail below.

Q1. Higher-Resolution Masks and Alternative Mask Objectives (1/2)

To address the reviewer's concerns regarding the effects of mask resolution and training objectives on segmentation performance, we conducted two additional experiments. First, we progressively increased the mask output and RoIAlign resolutions to examine the effect of spatial decoding capacity. Second, we applied boundary supervision corresponding to LiNA's Curvature-Aware Boundary (CAB) loss to dense mask decoders to investigate whether differences in training objectives account for the observed performance gap.

1. Higher-Resolution Mask Decoding

We first increased the output mask resolution from 28×28 to 56×56 while keeping the RoIAlign resolution fixed at 14×14. We then increased the RoIAlign resolution from 14×14 to 28×28 while maintaining the 56×56 output resolution.

All three dense decoders used the same backbone, detector, classifier, prototype distillation, and training configuration, with only the mask decoder resolution settings varied.

Table Q1-1. Effect of mask output and RoIAlign resolution on segmentation performance and computational cost.

ModelMask ResolutionRoIAlignDet-F1Cls-F1avgPQFμ (G)Fmax (G)
LiNA (Ours)Fourier contour—88.0871.9145.1248.048.0
Dense2828×2814×1487.3870.4744.0685.6307.4
Dense56-A56×5614×1487.4670.3444.22125.0620.0
Dense56-B56×5628×2887.4270.3744.36181.81,069.1

Results are averaged over three PanNuke folds. Fμ and Fmax follow the computational-cost definitions in Table 5(b), corresponding to 24 and 190 decoded RoIs, respectively.

Increasing the output mask resolution from 28×28 to 56×56 improved avgPQ from 44.06 to 44.22 (+0.16 percentage points). However, Fμ increased from 85.6 to 125.0 GFLOPs, while Fmax nearly doubled from 307.4 to 620.0 GFLOPs.

Further increasing the RoIAlign resolution from 14×14 to 28×28 yielded only another 0.14 percentage-point improvement in avgPQ (44.22 to 44.36). In contrast, Fμ increased from 125.0 to 181.8 GFLOPs, and Fmax from 620.0 to 1,069.1 GFLOPs. Neither Det-F1 nor Cls-F1 showed consistent improvements with increasing resolution.

In comparison, LiNA achieved an avgPQ of 45.12 at 48.0 GFLOPs, outperforming the highest-resolution Dense56-B decoder while requiring approximately 3.8× fewer FLOPs at Fμ and 22.3× fewer FLOPs at Fmax.

These results demonstrate that substantially increasing the output resolution and spatial decoding capacity of dense mask decoders yields only limited improvements in segmentation performance. This supports our central observation that high-capacity pixel-wise shape decoding is not necessarily required to achieve strong nuclei segmentation accuracy.

Q1. Higher-Resolution Masks and Alternative Mask Objectives (2/2)

2. Alternative Mask Objectives

We further investigated whether differences in training objectives between LiNA and the dense decoders could explain the observed performance gap. LiNA employs Fourier contour supervision together with CAB loss, whereas the original dense decoders are trained using binary cross-entropy (BCE) and soft Dice losses.

To examine this possibility, we retained the original BCE and Dice losses and introduced two additional boundary supervision terms corresponding to CAB.

The first is a Signed Distance Field (SDF) loss that encourages alignment between predicted and ground-truth boundaries. The second is a curvature-landmark loss that encourages accurate prediction of high-curvature locations on the ground-truth boundary.

While these objectives are applied directly to explicit boundary points in the Fourier contour representation, we adapted them to dense masks using the boundary strength of the predicted soft mask and its 0.5 probability contour. Both terms are computed from continuous predictions, allowing gradients to propagate to the mask logits. We used the same loss weights and activation schedule as in LiNA (SDF: 0.5; landmark: 0.3), while keeping the dense decoder architecture and all other training conditions unchanged.

Table Q1-2. Effect of CAB-equivalent boundary supervision on dense mask decoders.

ModelTraining ObjectiveDet-F1Cls-F1avgPQ
LiNA (Ours)Fourier contour loss + CAB88.0871.9145.12
Dense28BCE + Dice87.3870.4744.06
Dense28 + CABBCE + Dice + CAB-equivalent87.4370.3044.05
Dense56-ABCE + Dice87.4670.3444.22
Dense56-A + CABBCE + Dice + CAB-equivalent87.5970.5244.39
Dense56-BBCE + Dice87.4270.3744.36
Dense56-B + CABBCE + Dice + CAB-equivalent87.5770.5044.22

Results are averaged over three PanNuke folds. The effect of boundary supervision is evaluated by comparing each dense decoder with its CAB-augmented counterpart under the same experimental protocol.

Adding CAB-equivalent boundary supervision did not consistently improve dense mask segmentation performance. For Dense28, avgPQ decreased slightly from 44.06 to 44.05 (−0.01 percentage points). Dense56-A improved from 44.22 to 44.39 (+0.17 percentage points), whereas the highest-resolution Dense56-B decreased from 44.36 to 44.22 (−0.14 percentage points).

Moreover, the direction of the avgPQ change varied across PanNuke splits for all three decoders. This indicates that additional boundary supervision can provide small improvements for certain decoder configurations or data splits, but does not consistently improve segmentation performance across dense mask decoders.

The effects on detection and classification were also limited. For Dense28, Det-F1 increased by 0.05 percentage points while Cls-F1 decreased by 0.17 percentage points. Dense56-A showed increases of 0.13 and 0.18 percentage points, respectively, while Dense56-B showed increases of 0.15 and 0.13 percentage points. Although Det-F1 improved slightly in all three configurations, these changes did not translate into consistent improvements in avgPQ.

One possible explanation lies in the different supervision characteristics of dense masks and Fourier contours. Dense decoders already receive pixel-level foreground/background supervision through BCE and optimize overall mask overlap through Dice loss. Consequently, the additional boundary information provided by SDF and curvature-landmark supervision may partially overlap with their existing pixel-level objectives.

In contrast, Fourier contours represent the entire nucleus boundary using a limited number of coefficients. The Fourier loss supervises the overall contour geometry by comparing reconstructed contours with ground-truth contours, but additional supervision can still improve alignment with the actual boundary and the preservation of local high-curvature details. CAB is designed to provide this complementary boundary supervision.

Importantly, none of the three dense decoders with CAB-equivalent supervision reached LiNA's avgPQ of 45.12. Even the best-performing configuration, Dense56-A + CAB, achieved 44.39, remaining 0.73 percentage points below LiNA. Furthermore, because CAB-equivalent supervision modifies only the training objective, the inference FLOPs of the dense decoders remain unchanged.

Taken together, these results show that applying boundary supervision corresponding to LiNA's CAB loss yields limited and inconsistent improvements for dense mask decoders. Thus, the performance difference between dense decoding and LiNA cannot be explained solely by the presence or absence of additional boundary supervision.

AuthorsOpenReview ↗

Q2. Statistical Uncertainty and Robustness Across Random Seeds

Q2. Statistical Uncertainty and Robustness Across Random Seeds

To address the reviewer's concerns regarding statistical uncertainty, we conducted additional multi-seed evaluations and bootstrap confidence interval analyses to assess the robustness of LiNA to training initialization and evaluation sample variability. These two analyses quantify distinct sources of uncertainty: variation across training seeds and uncertainty arising from the composition of test images.

1. Multi-Seed Evaluation

We evaluated each method using three random seeds across all three PanNuke folds. Table Q2-1 reports the mean performance and standard deviation (SD) across seeds.

Table Q2-1. Multi-seed performance on PanNuke (mean ± seed SD, %).

ModelDet-F1Cls-F1avgPQ
LiNA (Ours)88.09 ± 0.0471.93 ± 0.1645.13 ± 0.19
Dense2887.39 ± 0.0570.46 ± 0.2344.07 ± 0.23
Dense56-A87.47 ± 0.0470.53 ± 0.2944.45 ± 0.24
Dense56-B87.47 ± 0.1170.30 ± 0.3244.14 ± 0.31

LiNA achieved the highest mean Det-F1, Cls-F1, and avgPQ among all compared configurations. In particular, LiNA obtained an avgPQ of 45.13 ± 0.19, achieving both the highest mean performance and the lowest seed SD among the compared models. These results demonstrate that LiNA maintains stable joint nuclei analysis performance across different training initializations. Moreover, the avgPQ differences among the dense configurations (at most 0.38 points) are comparable to their seed SDs (0.23–0.31), whereas LiNA's lead over every dense configuration (at least 0.68 points) exceeds the seed SD of all models.

2. Bootstrap Confidence Intervals

Separately from training-seed variability, we assessed robustness to the composition of evaluation images. With the selected checkpoints fixed for each PanNuke split, we resampled test images with replacement 2,000 times and recomputed avgPQ using the official evaluation procedure for each resample.

For model comparisons, we employed a paired bootstrap using identical resampled images for both models. We calculated 95% confidence intervals (CIs) from the 2.5th and 97.5th percentiles of the bootstrap distribution of performance differences.

Table Q2-2. Image-level paired bootstrap results for avgPQ (2,000 resamples).

Model / ComparisonavgPQ / Difference95% CI
LiNA (Ours)45.12[44.46, 45.81]
Dense56-A − LiNA−0.89[−1.30, −0.49]
Dense56-B − LiNA−0.76[−1.18, −0.33]

LiNA outperformed Dense56-A and Dense56-B by 0.89 and 0.76 percentage points in avgPQ, respectively. Neither paired bootstrap 95% CI includes zero, indicating that the observed performance advantages remain robust to variations in the composition of evaluation images.

Importantly, this bootstrap analysis quantifies evaluation-sample uncertainty for fixed checkpoints, which is distinct from the training-initialization variability assessed in the multi-seed evaluation.

Taken together, the multi-seed evaluation and paired bootstrap analysis provide complementary evidence regarding robustness to training initialization and evaluation-sample variability. Both analyses support LiNA's performance advantage and strengthen the statistical evidence underlying our accuracy–efficiency comparison.

Reviewer_1hnqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes LiNA, an efficient framework for joint nuclei analysis that reduces the capacity required for nuclear geometry representation while preserving recognition performance. The method is motivated by the observation that nuclear shapes can be accurately represented using low-order geometric representations, avoiding the computational overhead of dense-mask decoding. Experiments on PanNuke, Lizard, MoNuSeg, and CPM17 demonstrate promising accuracy and efficiency, including zero-shot generalization across datasets.

Critical strengths

The paper explores an interesting direction of improving computational efficiency by reducing unnecessary representation capacity for nuclear geometry.

Critical weaknesses
  1. Although segmentation is identified as one of the main tasks and contributions, it is unclear how segmentation results are generated from the architecture in Figure 1 or how segmentation is incorporated into the loss function.

  2. The importance and contribution of each loss term in the overall objective are unclear. More explanation is needed to justify the design of the loss function.

  3. Many important implementation details of the proposed method are missing. For example, the fusion component illustrated in Figure 1 is not sufficiently explained, making it difficult to understand how different features are combined.

  4. The overall writing and presentation are confusing. The description of the proposed framework lacks clarity, making the methodology difficult to follow.

Questions for authors
  1. How are the segmentation masks generated by LiNA? Which component in Figure 1 is responsible for producing the segmentation predictions?

  2. How does the loss function supervise the segmentation task, and what is the relationship between the geometric representation and segmentation output?

  3. How are the loss weights determined?

  4. How does the fusion component in Figure 1 operate, and which features are fused?

  5. Can the authors clarify the overall workflow of LiNA and the connections between its major components?

7 comments
AuthorsOpenReview ↗

Q1. Segmentation Mask Generation and the Responsible Module

We sincerely thank the reviewer for the careful evaluation of our manuscript and the constructive feedback. We particularly appreciate the comments on the clarity of our architecture, segmentation procedure, and training objectives. These insightful suggestions have helped us substantially improve the clarity and accessibility of our presentation. In response, we have redesigned Figure 1 and refined the methodological descriptions to make the overall framework and its components more intuitive and easier to follow. We address each question in detail below.

Q1. Segmentation Mask Generation and the Responsible Module

We have added Figure 2 (page 4 of the revised PDF), a detailed view of LiNA, to explicitly illustrate how LiNA generates instance segmentation masks and which components are responsible for this process. Figure 1 remains the overview of the training and inference stages.

As shown in Figure 2(a), the Detection Head first predicts nucleus centers, and the ROI feature sampled at each center from the student feature map is passed to the Geometry Path. There, the Fourier Head predicts the real and imaginary parts of 2N+1 = 31 complex Fourier coefficients (62 real values, N = 15) for each nucleus. These coefficients are decoded into a closed contour of 128 points, which is then rasterized into the final binary instance mask.

Thus, the Fourier Head is the module that predicts nucleus geometry, while rasterization is a deterministic operation that converts the reconstructed contour into an instance segmentation mask. This process does not require a separate learnable dense mask decoder.

Figure 2 shows the complete sequence of Fourier coefficient prediction → contour reconstruction → rasterization → instance mask, while clearly distinguishing the parallel Geometry and Recognition Paths.

AuthorsOpenReview ↗

Q2–Q3. Segmentation Supervision, Loss Contributions, and Weight Selection

Q2–Q3. Segmentation Supervision, Loss Contributions, and Weight Selection

We clarify how LiNA supervises segmentation, the roles and contributions of each loss component, and the criteria used to select loss weights. The newly added Figure 2 illustrates the corresponding components and training procedures.

1. Task-Specific Supervision

LiNA employs three primary loss functions:

  • Detection Loss (L_det): Smooth L1 loss is applied to matched nucleus centers obtained through Hungarian matching to supervise localization.
  • Classification Loss (L_cls): Binary cross-entropy loss supervises nucleus type predictions.
  • Fourier Loss (L_fourier): A cyclic-shift-invariant Smooth L1 loss between reconstructed nucleus contours provides direct supervision for segmentation.

In particular, the Fourier loss serves as the primary supervision signal for segmentation. LiNA extracts nucleus boundaries from ground-truth instance masks and represents them as 128-point contours. Similarly, the Fourier coefficients predicted by the Fourier Head are reconstructed into 128-point contours. A cyclic-shift-invariant Smooth L1 loss is applied between these contours, allowing the model to learn nucleus geometry regardless of the starting-point ambiguity of closed contours.

Importantly, the loss is applied directly to the reconstructed contours, rather than to the rasterized binary masks. Since Fourier reconstruction is differentiable, gradients propagate through the Fourier coefficients to the Fourier Head. The predicted contours are subsequently rasterized to generate the final instance segmentation masks.

2. Additional Supervision and Weight Selection

Curvature-Aware Boundary (CAB) Loss (L_CAB) complements Fourier supervision by aligning predicted contours with ground-truth boundaries through a signed-distance field and preserving local geometric details through curvature-landmark supervision. We compared CAB loss weights of 0.5 and 1.0 and selected 0.5 based on validation avgPQ.

Prototype Distillation Loss (L_proto) is applied through a two-phase training process. In Phase 1, teacher-derived prototype embeddings are concatenated with student RoI features, and the codebook is learned through the classification loss. In Phase 2, the codebook is frozen, and a foreground-masked KL divergence aligns the student and teacher soft code distributions (Figure 2(c)). At inference, only the student and the learned codebook are used, without the teacher. We explored loss weights of 0.1, 1.0, and 2.0 and selected 2.0 based on the magnitude of the distillation loss relative to the primary task losses.

3. Quantitative Contributions

ConfigurationCls-F1avgPQ
Full LiNA0.7190.451
w/o CAB0.7090.440
w/o Prototype Distillation0.6400.436

Results on PanNuke. The "w/o Prototype Distillation" configuration removes the entire codebook-based prototype pathway.

Removing CAB reduces the final joint nuclei analysis performance, demonstrating the contribution of boundary supervision beyond the primary Fourier loss. Removing the Prototype Distillation pathway leads to a particularly large decrease in Cls-F1, highlighting the importance of prototype representations that transfer semantic knowledge from the teacher.

AuthorsOpenReview ↗

Q4. Clarification of Feature Fusion (Concatenation)

Q4. Clarification of Feature Fusion (Concatenation)

We acknowledge that the ambiguity surrounding "Fusion" may have arisen from how Prototype Distillation and Fusion were presented as separate components in the original manuscript. In particular, the title "Prototype Distillation and Fusion" in Section 3.2 and the "Fusion" label in the original Figure 1 could have suggested that feature concatenation was an independent module or an additional learnable mechanism.

In LiNA, this operation is simply the concatenation of the student RoI feature r_i, extracted at each detected nucleus center, and the codebook-derived prototype embedding e_i = z_i E along the feature dimension. The concatenated feature f̂_i = [r_i ‖ e_i] is then passed to the Classification Head. The concatenation itself introduces no learnable parameters. This operation was already defined in Section 3.2 of the original submission, following Eq. (2).

To eliminate this ambiguity and more accurately reflect the implementation, we have revised the relevant terminology and presentation throughout the manuscript:

  • Figures 1 and 2: We replaced the "Fusion" label in Figure 1 with "Concat", and the newly added Figure 2(a) explicitly illustrates how the student RoI feature and prototype embedding are concatenated before being passed to the Classification Head.
  • Section 3.2: We renamed "Prototype Distillation and Fusion" to "Prototype Distillation," clarifying that feature concatenation is not an independent module.
  • Table 6 (Ablation Study): We changed "+ Proto-Distill and Fusion" to "+ Proto-Distill" to avoid implying separate contributions. This configuration introduces the complete codebook-based Prototype Distillation pathway; the experimental setup and results remain unchanged.
  • Text: In Sections 3.2 and 4, "fuse" and "fused" now read "concatenate" and "concatenated" (or "use"), and "proto-distill and fusion" now reads "proto-distill".

These revisions clarify that Prototype Distillation is a unified mechanism for transferring semantic knowledge, while concatenation is simply the operation used to incorporate the resulting prototype embeddings into classification. The new Figure 2(a) and Section 3.2 now explicitly show the sources of the features, how they are combined, and how the concatenated representation is passed to the Classification Head.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.