acceptodds
Under review as a conference paper at ICLR 2027

Multi-Frequency Consistency and Spectral-Basis Caching for Spectral-Aware Staining

Abstract

Virtual staining enables the generation of biologically informative microscopy images without physical staining procedures. However, conventional evaluation and training objectives may overlook fine-grained biological structures. Downstream evaluation is more biologically relevant but requires task-specific annotations, predictive models, or expert assessment, limiting the scalability of downstream evaluation. Meanwhile, diffusion-based virtual staining relies on iterative denoising and can incur substantial inference cost. We propose Spectral-Aware Staining, a spectral-guided framework for virtual staining that improves staining quality assessment and diffusion acceleration. First, we introduce Multi-Frequency Spectral-Aware Consistency (MSAC), a spectral-aware consistency metric that evaluates virtual staining quality by comparing retained multi-frequency spectral subspaces. By enforcing multi-frequency consistency, the MSAC better preserves biologically relevant morphology. Second, we propose Spectral-Basis Cache (SBCache), a structure-aware caching strategy for diffusion inference. SBCache estimates denoising redundancy in a learned graph-spectral space and enables reuse of intermediate computations while achieving a favorable quality–efficiency trade-off. Both MSAC and SBCache share the same trained neural spectral decomposer, which predicts ordered graph-spectral bases from feature graphs and replaces computationally expensive eigendecomposition with efficient network inference. Experiments on public microscopy benchmarks and virtual-staining frameworks demonstrate that MSAC more effectively evaluates the quality of generated stains and that Spectral-Aware Staining generates high-quality stained images more efficiently.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Virtual Staining, Multi-Frequency Spectral Analysis, Graph Spectral Representation, Image Quality Assessment, Microscopy Image Synthesis, Diffusion Acceleration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_VXnBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Spectral-Aware Staining combines a graph-spectral consistency measure with diffusion caching. A neural decomposer learns from offline numerical eigenvectors using subspace alignment, Rayleigh off-diagonal separation, and spectral ordering. MSAC compares retained subspaces and can supervise staining models, while SBCache filters timestep-dependent denoiser features before measuring redundancy. Experiments on cpg0000-jump-pilot and HEMIT include five backbones, controlled degradations, CellProfiler segmentation/counting, encoder ablations, repeated well-level splits, and matched-refresh cache comparisons. MSAC improves all 25 reported downstream comparisons. SBCache generally improves the cache trade-off, but its shared-decomposer feature-space transfer, independent biological validity, and total preparation cost need further evidence.

Critical strengths
  1. Structural improvements are consistent under substantial experimental controls. MSAC improves all 25 reported segmentation and counting comparisons across five backbones and two datasets. Matched additional optimization, perceptual-loss baselines, controlled degradations, and five well-level splits strengthen this evidence. These results support acceptance because the loss offers practical benefits beyond improving its own score, consistently across architectures rather than only one favorable implementation.

  2. The spectral representation is technically motivated and experimentally supported. Subspace alignment, Rayleigh separation, and ordering address sign ambiguity, mixed directions, and frequency identity. Image-feature decomposition accuracy exceeds SBLNet and SpecNet2, while MSAC correlates with CellProfiler structural quality on predictions produced without MSAC supervision. This provides evidence for a useful learned representation beyond introducing another perceptual distance, strengthening the methodological case for acceptance.

Critical weaknesses
  1. The shared decomposer is not validated on the actual cache-feature domain. It is trained on compact image-encoder features but receives JiT-UNet features directly in SBCache, without reported alignment or adaptation. Matching dimensions does not establish compatible channel semantics or spectral accuracy. A central claim of the unified framework depends on this transfer, yet no timestep-specific eigenspace or ordering diagnostics verify it. This unresolved assumption materially weakens the case for acceptance.

  2. Independent biological validation remains insufficient. The downstream masks and counts come from the same CellProfiler pipeline on real and synthesized stains, rather than independent annotations. Controlled degradations and correlations on models trained without MSAC are useful safeguards, but do not establish agreement with expert-defined biological quality. Because biological structure preservation is the central motivation, dependence on one automated proxy limits the strength of the utility claim.

Questions for authors
  1. Why was cpg0000-jump-pilot selected from the Cell Painting Gallery, and how were the 4,700 images sampled? Please clarify the coverage of cell lines, perturbations, plates, and batches. Could the main gains be validated on another Gallery dataset with paired brightfield–fluorescence images, or on a held-out acquisition batch, to assess generalization beyond the current well-level splits?

  2. Can the shared decomposer be evaluated against numerical eigentargets from actual JiT-UNet features at different timesteps, reporting subspace error, Rayleigh-eigenvalue error, and ordering violations? How does alignment or separate training affect results?

  3. How sensitive are the reported structural improvements to the choice of segmentation pipeline? If feasible, an additional check using an existing alternative segmentation method would strengthen the evidence. Otherwise, please clarify the limitations of interpreting CellProfiler-derived mask agreement as biological fidelity.

  4. Can the authors report segmentation/counting after caching and additional diffusion backbones, explain HEMIT intensity-versus-structure trade-offs, and test whether ordering losses improve final utility beyond decomposer diagnostics?

Reviewer_oKR8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spectral-Aware Staining to address the inaccurate evaluation of fine-grained biological structures in virtual staining and the low inference efficiency of diffusion models. Building on a Neural Spectral Decomposer that learns ordered graph spectral bases, the method introduces Multi-Frequency Spectral-Aware Consistency (MSAC) to enhance structural fidelity and Spectral-Basis Cache (SBCache) to accelerate diffusion inference.

Critical strengths
  1. The research problem is clearly defined. The paper demonstrates the discrepancy between pixel-level similarity and biological structural fidelity, highlighting that the ability of generated images to support downstream biological analysis better reflects the practical requirements of virtual staining.

  2. The graph frequency modeling exhibits a certain degree of novelty. The authors further employ a neural network to learn spectral bases and introduce subspace alignment and frequency-ordering constraints to address eigenvector sign ambiguity and permutation issues. The method design has a mathematical foundation.

  3. MSAC exhibits good modularity and extensibility. It can not only measure structural differences between generated and ground-truth images but also be incorporated into different models as a differentiable loss function.

  4. The experiments are relatively comprehensive, covering various settings and models. The experiments also evaluate the effectiveness of MSAC as an evaluation metric.

Critical weaknesses
  1. The benefit of the Neural Spectral Decomposer for MSAC remains unclear. MSAC employs a rotation-invariant subspace projection distance. Even though the Neural Spectral Decomposer learns ordered spectral bases, does MSAC itself actually utilize the specific frequency ordering of these bases?

  2. The effectiveness and generalizability of SBCache have not been sufficiently demonstrated. Although SBCache is presented as a plug-and-play module, its applicability across different model architectures is not adequately evaluated. Compared with SeaCache, its relative latency improvement is limited, and some metrics on the HEMIT dataset are inferior to those of SeaCache.

  3. The shared Spectral Decomposer may suffer from feature distribution mismatch. MSAC and SBCache share the same trained Neural Spectral Decomposer, but the authors do not explain why a Spectral Decomposer trained on stained image features can accurately process the intermediate features of diffusion models.

Questions for authors
  1. The benefit of the Neural Spectral Decomposer for MSAC remains unclear. MSAC employs a rotation-invariant subspace projection distance. Even though the Neural Spectral Decomposer learns ordered spectral bases, does MSAC itself actually utilize the specific frequency ordering of these bases?

  2. The effectiveness and generalizability of SBCache have not been sufficiently demonstrated. Although SBCache is presented as a plug-and-play module, its applicability across different model architectures is not adequately evaluated. Compared with SeaCache, its relative latency improvement is limited, and some metrics on the HEMIT dataset are inferior to those of SeaCache.

  3. The shared Spectral Decomposer may suffer from feature distribution mismatch. MSAC and SBCache share the same trained Neural Spectral Decomposer, but the authors do not explain why a Spectral Decomposer trained on stained image features can accurately process the intermediate features of diffusion models.

Reviewer_ZTHeOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Spectral-Aware Staining, a framework for improving structural fidelity assessment and inference efficiency in virtual staining. The framework consists of two main components. First, Multi-Frequency Spectral-Aware Consistency (MSAC) measures structural similarity between generated and reference staining images by comparing their learned graph-spectral subspaces. MSAC can be used both as an evaluation metric and as an additional training objective. Second, Spectral-Basis Cache (SBCache) exploits graph-spectral representations of intermediate diffusion features to guide caching decisions and accelerate inference. Both components share a neural spectral decomposer trained to approximate numerical graph eigendecomposition.

The authors evaluate MSAC on two virtual-staining datasets, cpg0000-jump-pilot and HEMIT, using five generative architectures. The results demonstrate consistent improvements in downstream segmentation accuracy and cell counting, and show that MSAC correlates more strongly with downstream segmentation performance than conventional image-quality metrics and direct perceptual feature distance. Controlled degradation experiments further support its sensitivity to small structural changes. For diffusion acceleration, SBCache achieves favorable quality–efficiency trade-offs compared with existing caching methods, particularly on cpg0000-jump-pilot.

Critical strengths
  1. Well-motivated and technically interesting methodology.

The proposed graph-spectral approach is a thoughtful methodological development for virtual staining. The motivation to move beyond conventional pixel-level similarity and capture structural relationships in learned image features is reasonable, particularly for preserving fine-grained biological morphology. I also appreciate the unified design, where a learned neural spectral decomposer enables both structural consistency assessment (MSAC) and efficient diffusion inference (SBCache). Although the biological interpretation of the graph-spectral representation could be better demonstrated, I find the overall methodology technically sound and potentially valuable.

  1. Consistent improvements in biologically relevant downstream tasks.

MSAC demonstrates consistent improvements in downstream segmentation (Dice and IoU) across five virtual-staining architectures and two datasets, as well as improved cell-counting accuracy on JUMP (Table 1). Some improvements are substantial, particularly when conventional image-quality metrics show only marginal gains or even deteriorate. I consider this an important strength because the preservation of biologically meaningful structures, rather than pixel-level similarity alone, is a central objective of virtual staining.

  1. Complementary experiments supporting the value of MSAC.

The authors provide several well-designed experiments supporting the proposed metric. Figure 4 demonstrates that MSAC correlates strongly with downstream segmentation performance, while Table 2 shows greater sensitivity to small-structure removal and boundary degradation than to Gaussian noise at matched PSNR. Importantly, the ablation studies in Tables 4 and 8 demonstrate improvements over direct pretrained-feature matching across different feature encoders. Together, these results provide encouraging evidence that MSAC offers additional value beyond conventional image-quality and perceptual feature-matching objectives.

Critical weaknesses
  1. Insufficient explanation and demonstration of the biological relevance of the graph-spectral representation.

Although I find the proposed methodology interesting and the quantitative results encouraging, the manuscript could substantially improve how it communicates the central scientific motivation and contribution of MSAC. In particular, it remains unclear what biologically meaningful information the learned graph-spectral representation captures and how this relates to the preservation of subcellular structures in virtual staining. For example, do the graph affinities capture relationships between mitochondrial structures or other intracellular features, and how does MSAC exploit these relationships to improve reconstruction? A visualization connecting the graph representation to actual subcellular morphology would help readers understand the specific advantages of this methodological approach.

This issue also extends to the qualitative evaluation in Figure 5. Although Table 1 shows convincing improvements in downstream segmentation and counting, the authors do not clearly demonstrate which structures are recovered and how these improvements translate into more accurate biological measurements. Enlarged regions, arrows highlighting corrected structures, and corresponding segmentation overlays would make the results much more interpretable. In particular, the substantial changes in the HEMIT CD3 predictions deserve closer examination to distinguish correctly recovered signals from potential false positives. Channel-specific Dice and IoU results would further clarify which biological structures benefit most from MSAC.

I consider this important because the current manuscript provides encouraging evidence that MSAC works, but does not yet communicate sufficiently clearly why the proposed graph-spectral representation is particularly suitable for this biological problem.

  1. Limited assessment of cross-plate generalization on JUMP.

The authors use a well-level train/test split for the JUMP dataset, which appropriately avoids overlap between images from the same well. However, this does not establish robustness to plate-associated batch effects, which are particularly relevant in high-content microscopy. A held-out-plate evaluation would provide stronger evidence that the improvements from MSAC generalize beyond closely related imaging conditions. At minimum, the authors should clarify how many independent plates and batches are represented in each split and discuss this limitation.

Questions for authors
  1. Biological interpretation of the graph-spectral representation.

Could the authors clarify what biologically relevant subcellular information is captured by the proposed graph-spectral representation, and how this contributes to improved virtual staining? For example, a visualization connecting graph affinities or spectral representations to mitochondrial or other subcellular structures in JUMP would help explain the underlying mechanism and the specific advantage of MSAC. It would also be helpful to connect the qualitative improvements in Figure 5 to downstream segmentation and counting results, ideally through highlighted regions, segmentation overlays, and channel-specific Dice/IoU measurements.

  1. Cross-plate generalization on JUMP.

The current JUMP evaluation uses a well-level train/test split. Could the authors clarify how many plates and experimental batches are represented and whether the splits share plates? More importantly, would the improvements from MSAC persist under a held-out-plate evaluation? Such an experiment would provide stronger evidence of robustness to plate-associated batch effects, which are particularly relevant for high-content microscopy.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.