acceptodds
Under review as a conference paper at ICLR 2027

PV-WSI: Rethinking Pseudo-Video Learning for Whole Slide Image Segmentation via Spatial Discontinuity Modeling

Abstract

Whole-slide image (WSI) segmentation for computational pathology is challenged by gigapixel scale, extreme morphological heterogeneity, and fragmented cross-patch context. Recent pseudo-video formulations attempt to treat WSI patches as sequential frames, but inherit a critical flaw: they impose temporal smoothness priors from natural video onto inherently discontinuous pathological landscapes. This mismatch leads to erroneous context propagation and degraded segmentation, especially across structurally dissimilar regions. We introduce PV-WSI, a pseudo-video paradigm that explicitly models both spatial continuity and discontinuity. Unlike prior work that passively adapts video architectures, PV-WSI reconstructs pseudo-temporal structure. Specifically, we propose Adaptive WSI Pseudo-Video Construction, which identifies pathology-informative anchor patches and dynamically assembles morphology-consistent sequences, eliminating traversal bias and enforcing structural alignment. Furthermore, we develop Distance-Constrained Gate State Propagation, a principled distance-aware propagation mechanism that selectively transmits reliable contextual states while suppressing spurious long-range dependencies across weakly correlated regions. Extensive experiments show that PV-WSI consistently outperforms state-of-the-art methods across magnification levels, improving both tumor segmentation and multi-class tissue parsing. These results demonstrate that explicitly modeling spatial discontinuity is essential for pseudo-video learning and support PV-WSI as a strong framework for large-scale WSI understanding. Code will be available at GithubGithub.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
whole-slide image segmentation;computational pathology;pseudo-video learning;medical image analysis

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_jN42OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PV-WSI, a pseudo-video framework for whole-slide image segmentation that explicitly addresses spatial discontinuities among pathology patches. The authors argue that directly applying video-style temporal propagation to WSIs can introduce unreliable contextual information because spatially neighboring patches may exhibit substantially different tissue morphologies. To address this issue, the proposed framework incorporates two main components. Adaptive WSI Pseudo-Video Construction (AWPVC) reorganizes patches into morphology-consistent sequences using importance-based initialization, structural similarity, and adaptive restarting. Distance-Constrained Gate State Propagation (DCGSP) selectively regulates historical information through spatial-distance and semantic-consistency cues. Experiments on CAMELYON16 and BCSS demonstrate improvements over several segmentation baselines. The paper further provides component ablations, comparisons with alternative context operators, transition perturbation experiments, backbone portability analysis, and computational efficiency measurements.

Critical strengths

The paper identifies a meaningful limitation of applying temporal propagation mechanisms to histopathology images. Unlike natural video frames, spatially adjacent WSI patches do not necessarily exhibit smooth structural transitions. The proposed combination of morphology-aware sequence construction and selective state propagation offers a coherent approach to addressing this issue. This is potentially useful for sequential modeling of large pathology images. The reported improvements are substantial across multiple magnifications on CAMELYON16 and across tissue categories on BCSS. The same-backbone ablations in Tables 3, 5, and 6 provide evidence that both proposed components contribute to performance. Moreover, the alternative context-operator comparison (Table 7), MedSAM2 portability experiment (Table 9), and transition perturbation analysis (Table 11) strengthen the empirical motivation beyond a simple comparison against existing methods. These experiments constitute an important strength of the submission.

Critical weaknesses
  1. Limited evaluation scope and insufficient evidence of statistical robustness. Although the reported performance improvements are promising, the experimental evidence does not fully establish the broader applicability and reliability of the proposed framework. First, the CAMELYON16 evaluation includes only annotated tumor-positive WSIs, while the BCSS evaluation is restricted to annotated ROIs rather than complete whole-slide inference. Both datasets concern breast histopathology. Consequently, the experiments do not establish generalization to other organs, medical centers, staining conditions, or normal-slide screening scenarios. Second, Appendix A explicitly states that no validation set was used and that repeated-seed experiments, confidence intervals, and significance tests are unavailable. While the authors state that test data were excluded from model selection, the absence of independent validation and uncertainty estimates makes it difficult to assess the stability of the reported gains and the robustness of hyperparameter choices. These limitations are important because the central empirical claim concerns the effectiveness of a general discontinuity-aware pseudo-video modeling paradigm rather than performance on a single narrowly defined benchmark setting.

  2. Insufficient specification and inconsistencies in the proposed algorithm. Several aspects of the method require clarification to establish its precise behavior and computational properties. First, Eq. (4) performs adaptive restarting through an optimization over the remaining unvisited patches, with a diversity term defined relative to previously visited frames. However, Appendix B claims linear sequence-construction complexity, O(N), mainly by considering comparisons with at most four spatial neighbors. It is unclear how the global restart search and history-dependent diversity computation are implemented without additional computational complexity. Although the measured runtime scales approximately linearly over the tested configurations, this observation does not establish the claimed algorithmic complexity. Second, the definition of the parameter lambda is inconsistent. Section 3.1 defines lambda as the coefficient balancing patch importance and diversity during adaptive restarting. However, the discussion of Table 13 describes lambda as controlling historical state propagation and interprets lambda = 1 as unrestricted propagation. This interpretation does not follow from Eq. (4) or the propagation equations in Section 3.2. Finally, the precise aggregation used for the history-dependent diversity term is insufficiently specified. These issues make it difficult to reproduce the sequence construction procedure and independently verify its theoretical properties.

3.Based on my experience with AI usage, I suspect that the image in your Figure 1 was directly generated by AI, because it is not a vector image compared to your method diagram.

Questions for authors

Q1. Complexity and implementation of adaptive restarting.

Could the authors provide the exact implementation or pseudocode for Eq. (4), including how the diversity term is aggregated over previously visited patches? How often does adaptive restarting occur in practice, and how is the reported O(N) complexity maintained when searching the remaining unvisited patch set? Clarifying the worst-case complexity would help assess the scalability claim.

Q2. Parameter definitions and reproducibility.

Could the authors clarify the discrepancy in the interpretation of lambda between Eq. (4) and the discussion of Table 13? Is lambda exclusively the importance-diversity balancing parameter, or does the implementation contain an additional propagation-control parameter? Please also clarify the precise diversity computation and the hyperparameter settings used by the distance-aware gate.

Q3. Model selection and robustness of experimental conclusions.

Given that no validation set was used, how were lambda, beta, tau, and other architectural hyperparameters determined without consulting test performance? Could the authors provide repeated-seed results or slide-level uncertainty estimates for the principal comparisons? Such evidence would materially strengthen confidence that the reported gains are stable rather than dependent on particular training runs or model-selection choices.

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

The paper argues that pseudo-video WSI segmentation inherits a temporal-smoothness prior from natural video that does not hold across tissue boundaries, and proposes PV-WSI: (i) AWPVC, which orders patches by pathology-informed importance, greedy SSIM traversal among four spatial neighbours, and diversity-based restart; (ii) DCGSP, which gates a cross-attended memory by an MLP over inter-patch displacement and semantic agreement, scaled by a Gaussian distance prior. On CAMELYON16 at five magnifications it reports 89.95 mean Dice vs. 82.95 for the strongest baseline, and 84.13 mDice / 72.90 mIoU on BCSS, with same-backbone ablations, a context-operator swap, a MedSAM2 transfer test, and a transition-perturbation test.

Critical strengths
  1. The paper isolates the suspected cause instead of only adding capacity.
  2. The evaluation is a genuinely controlled comparison, and its arithmetic is auditable.
Critical weaknesses
  1. The gap to prior pseudo-video methods is carried by a backbone that the paper never specifies.
  2. The deployed operating point looks test-optimal, and there is no variance estimate.
Questions for authors

Q1. Please give the encoder, decoder and parameter counts, and state whether the "H&E branch" baseline (78.50 M) was re-trained by you or quoted from [18]. This decides whether the headline gap reflects your modules or your backbone.

Q2. With no validation split, what objective chose λ, β, τ, the restart rule and the checkpoint, and were Tables 12–13 run before or after freezing that configuration?

Q3. Define Table 8: the unit of a "transition", how "region" is defined, and whether counts are length-normalised. As printed, 6.4/2.4/5.3 cannot be interpreted, yet the claimed 2.5× reduction rests on them.

Q4. Is there evidence that Path-SAM2/PathSeqSAM fail because of the smoothness prior, e.g. error conditioned on transition type? Table 11 perturbs only your model, and your margin over MedSAM2 already falls from 19.75 Dice at 1.25× to 2.17 at 20×.

Q5. What are the parameter counts of each row in Table 3? Since AWPVC is parameter-free while DCGSP adds capacity, please report a capacity-matched DCGSP variant so gating can be separated from raw parameters.

Reviewer_d7PvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes PV-WSI, a pseudo-video segmentation framework for whole slide images that addresses spatial discontinuities across patch transitions. The method consists of two main components: Adaptive WSI Pseudo-Video Construction, abbreviated as AWPVC, and Distance-Constrained Gate State Propagation, abbreviated as DCGSP. AWPVC starts a sequence at the patch with the highest sum of edge density and intensity contrast, then greedily traverses four-connected neighboring patches based on structural similarity, falling back to a color-histogram diversity restart when local similarity drops below a threshold. DCGSP regulates recurrent hidden-state propagation between consecutive pseudo-frames using cross-attention, a learned gate conditioned on coordinate displacement and feature cosine similarity, a Gaussian distance prior, a Temporal State Space module, and shifted-window Spatially Selective Renewal. Experiments are conducted on CAMELYON16 across five magnifications and on BCSS for region-of-interest five-class tissue segmentation.

Critical strengths

The motivation of addressing spatial discontinuity when converting two-dimensional pathology patches into one-dimensional pseudo-video sequences is well grounded in the structural properties of histopathological slides. Demonstrating in Appendix Table 9 that the proposed traversal module can be plugged into an existing video segmentation pipeline like MedSAM2 to improve accuracy across magnifications provides useful evidence that patch ordering matters for sequential modeling.

Critical weaknesses
  1. The manuscript exhibits severe traces of automated large language model rewriting, resulting in poor readability, disjointed coherence, and verbatim text duplication. While the AI Use Statement admits using generative AI for language editing and manuscript reorganization, the actual text reads as if an LLM was prompted to insert defensive rebuttal disclaimers directly into the body of the paper. For instance, in the Introduction on page 2, line 082, the text abruptly inserts a defensive sentence stating that its primary contribution is not importance scoring, SSIM traversal, or distance gating in isolation, before those terms are even explained. Similarly, Related Work on page 3, line 134 abruptly concedes that distance-aware weighting is not itself unique, and the tables and appendix repeatedly insert robotic disclaimers stating that differences should be interpreted as descriptive point estimates. Worse still, in Appendix A on page 14, the entire paragraph spanning lines 721 to 733 is duplicated almost word for word in lines 744 to 755, repeating the exact same sentences about training on one NVIDIA A800 for 50 epochs, having no validation set, and lacking confidence intervals. Submitting a manuscript with copy-pasted paragraphs and stitched-together AI disclaimers reflects a serious lack of basic proofreading and makes the paper difficult to read.

  2. Core architectural specifications are missing, and the mathematical descriptions contain direct internal contradictions. First, the paper never specifies the exact network architecture of the image encoder, the decoder, or the Temporal State Space module in Equation 12, making it impossible to know what the 18.10 million parameter model in Table 4 or the shared backbone in Table 3 actually is. Second, the hyperparameter lambda is defined in Equation 4 as the weight balancing patch importance and color diversity during sequence restart in AWPVC. However, in Appendix B on page 18, lines 939 to 945, the authors analyze lambda as a historical state propagation control parameter where setting lambda to one corresponds to unrestricted historical propagation. In reality, Equation 4 has nothing to do with state propagation, and the state propagation module in Equations 5 to 12 contains no such parameter at all, while the Gaussian distance scale parameter in Equation 10 is never defined or ablated anywhere in the paper. Third, Section 3.2 explicitly states on line 224 that using raw pixel coordinate distances causes numerical instability across gigapixel slides, yet Equation 9 feeds raw unnormalized coordinate offsets directly into the multilayer perceptron alongside the normalized distance. Moreover, during regular four-neighbor traversal in Equation 3, the stride-normalized distance in Equation 6 is always identically equal to one, meaning the Gaussian distance prior in Equation 10 is merely a constant scalar except when a sequence restart occurs.

  3. The experimental tables contain major data inversions and contradictions. In Table 1 on CAMELYON16, the Dice and mIoU columns are swapped across multiple methods and magnifications. Specifically, Table 1 reports PV-WSI achieving 81.69 percent Dice and 83.42 percent mIoU at 1.25x, 86.74 percent Dice and 88.67 percent mIoU at 2.5x, and 88.52 percent average Dice versus 89.95 percent average mIoU. By mathematical definition, IoU cannot exceed Dice, and these numbers directly contradict Tables 3 and 4 as well as line 410 in the main text, which report 83.42 percent Dice and 81.69 percent mIoU at 1.25x, and 88.67 percent Dice and 86.74 percent mIoU at 2.5x. The same impossible inversion where mIoU exceeds Dice appears for U-Net, Swin-UNet, SAM, Path-SAM2, MedSAM, MedSAM2, and H&E branch in Table 1. Furthermore, the efficiency analysis in Table 4 on page 9 contradicts Table 10 on page 17. Table 4 claims that at 1.25x magnification, AWPVC increases per-slide preprocessing time from 5.11 seconds to 8.86 seconds, which is a 3.75 second overhead, whereas Table 10 explicitly reports that AWPVC sequence construction at 1.25x takes only 0.56 seconds per slide. Table 4 also restricts the runtime comparison to 1.25x where each slide has only 21 patches, hiding the fact shown in Table 10 that at 20x magnification with 3,986 patches, AWPVC adds 108.2 seconds of preprocessing overhead per slide. Finally, no validation set is used in any experiment, and normal slides in CAMELYON16 are completely excluded.

Questions for authors
  1. Clarification request on architecture and mathematical definitions: What is the exact network architecture of the image encoder, the Temporal State Space module in Equation 12, and the decoder? How are the spatial feature dimensions pooled or aligned when computing the vector inner product in Equation 8 and concatenating it with scalar coordinate offsets in Equation 9, and why are unnormalized coordinate offsets fed into the gating network after Section 3.2 states that raw coordinates cause numerical instability?

  2. Clarification request on the hyperparameter conflict: Why does Appendix B on line 939 analyze lambda as a historical state propagation parameter when Equation 4 defines it solely as the importance-versus-diversity weight during sequence restart? What numerical value was used for the Gaussian variance parameter in Equation 10?

  3. Clarification request on experimental discrepancies: Why are the Dice and mIoU values inverted in Table 1 compared to Tables 3 and 4, and why does the AWPVC preprocessing time at 1.25x equal 3.75 seconds in Table 4 but 0.56 seconds in Table 10? What is the total end-to-end preprocessing and inference runtime of PV-WSI at 20x magnification compared to the baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.