acceptodds
Under review as a conference paper at ICLR 2027

PPFuse: Fuse What Aligns, Preserve What Does Not

Abstract

Frozen pathology foundation models provide complementary representations of the same tissue patch, motivating feature fusion for whole-slide image classification. Aligning a lower-dimensional auxiliary representation to an anchor subspace leaves the treatment of the anchor orthogonal complement unspecified. Full-space interpolation implicitly ties complement scaling to the reachable-space mixing weight, although the correspondence does not require this coupling. We introduce PPFuse, which fits a semi-orthogonal correspondence from paired training features without task labels and decouples these decisions. After scale-matched mixing in the reachable coordinates, PPFuse adopts a minimum-change principle relative to the anchor: for the fixed correspondence and fused coordinates, the unique closest reconstruction in Frobenius distance preserves the complement at its native scale. The resulting transform is closed-form and retains the anchor dimensionality. Across three multiclass whole-slide datasets and three multiple instance learning backbones, PPFuse achieves higher mean balanced accuracy than both individual encoders in eight of nine settings and ranks first among the evaluated methods in seven. Across five few-shot sampling seeds, PPFuse exceeds both individual encoders in mean ROC-AUC across all nine backbone-shot settings. Component analyses show that the retained complement carries predictive signal and support native-scale preservation as a competitive fixed default across the evaluated datasets and backbones. Alignment controls further show that downstream adaptation narrows the performance gap between fitted and random alignments, highlighting the distinction between geometric alignment and predictive utility. Additional natural-image and image-text experiments demonstrate competitive performance against the evaluated fusion baselines while retaining the anchor dimensionality.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, feature fusion, representation alignment, foundation models, computational pathology, multiple instance learning, multi-view learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Gd52OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

PPFuse proposes a closed-form feature fusion method for combining frozen pathology foundation model representations at the tile level. The method aligns an auxiliary representation to an anchor space using a semi-orthogonal Procrustes mapping, decomposes the anchor representation into an auxiliary-reachable component and an orthogonal complement, fuses only the reachable coordinates after scale matching, and preserves the orthogonal complement based on a minimum-change reconstruction principle.

Critical strengths

The method has a clear geometric rationale: it fuses aligned features while preserving the anchor information outside the aligned subspace. Its closed-form, label-free construction also makes it easy to use with existing MIL models.

Critical weaknesses
  1. The evaluation does not clearly isolate the value of the fused tile features. Pathology results depend on retrained MIL aggregators, the full-data and few-shot tasks differ, and CONCH v1.5 is evaluated using 256×256 crops rather than its recommended 512×512 setting. Tile-level tests and a native-crop control are needed to attribute the gains to the fusion method.
  2. The practical benefit of a second tile encoder is not established. The gains are modest, and all nine multiclass comparisons against UNI2-h have paired 95% confidence intervals that include zero. Without an end-to-end cost comparison or clearer positioning against existing multi-encoder methods such as ELF, the added extraction cost is difficult to justify.
Questions for authors
  1. Could the authors clarify why some tasks are evaluated only in the few-shot setting, while others use the full training set? Would the conclusions remain consistent if both settings were reported across all tasks?
  2. Could the authors clarify the patch size used for CONCH v1.5 and UNI2? If both use 256×256 crops, how do the authors account for the difference from the standard input setting of CONCH v1.5?
  3. Given that tile-level feature extraction is typically much more computationally expensive than slide-level aggregation, could the authors justify the practical benefit of fusing features from two tile encoders? The reported gains appear relatively modest compared with the additional cost of extracting both CONCH v1.5 and UNI2 features. In contrast, several WSI-level foundation models, such as TITAN, FEATHER, and CARE, are built on CONCH v1.5 features, making fusion at the WSI-representation level potentially much cheaper.
  4. Similar to Question 3, the authors use MIL to evaluate the effectiveness of tile-level feature fusion. I think this is only partially convincing, because the MIL aggregator is trained for each downstream task, meaning that the final performance is influenced not only by the fusion strategy but also by the aggregator itself. Moreover, different baselines may require different MIL hyperparameter settings for optimal performance. Therefore, evaluating the proposed fusion strategy on tile-level downstream tasks may provide a more direct and fair assessment of the quality of the fused tile representations.
  5. The paper does not compare against a closely related ensemble-based pathology foundation model approach, e.g., Xiyue Wang et al., “Ensemble Learning of Pathology Foundation Models for Precision Oncology”, which appears to be an important baseline for evaluating whether the proposed fusion strategy provides benefits beyond existing model-ensemble methods.

Overall, I am particularly interested in the authors’ responses to Questions 3 and 4, as they will substantially affect my final rating. I hope the authors can provide additional experiments to address these concerns. In particular, tile-level downstream tasks could be used to directly validate the effectiveness of tile-level feature fusion. It would also be valuable to explore fusion at the slide-representation level using WSI foundation models such as TITAN and CARE, to demonstrate whether the proposed fusion strategy remains effective across different representation levels rather than only under MIL-based pipelines. Finally, reporting the computational cost and runtime for the different fusion settings would help clarify the practical trade-offs. I believe these additional experiments would substantially strengthen the paper.

1 comment

The authors have still not addressed my primary concern. PPFuse essentially fuses representations from two tile-level pathology foundation models. Therefore, evaluation on a tile-level pathology benchmark (SICAPv2) would provide the most direct evidence of the effectiveness of the proposed method. However, the observed improvements on this benchmark are not particularly significant.

On the other hand, the slide-level feature fusion (TITAN + FEATHER) appears to result in a performance degradation. It remains unclear whether this is because the original FEATHER representations negatively affect the fused representations or due to other factors. The authors could consider evaluating the TITAN + CARE combination to further demonstrate the effectiveness of PPFuse.

Furthermore, ICLR allows authors to update their manuscripts during the rebuttal period. I believe the authors should incorporate the tile-level PPFuse results on SICAPv2 and the slide-level PPFuse results using TITAN + CARE into the main manuscript. Evaluating tile-level fusion on tile-level tasks and slide-level fusion on slide-level tasks provides the most direct evidence of the advantages of the proposed fusion strategy. In contrast, evaluating tile-level fusion combined with MIL on slide-level tasks serves only as supplementary evidence of the effectiveness of the fusion strategy. Including these results would help improve the paper's Informativeness and Soundness.

Reviewer_ADysOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PPFuse, a training-free way to fuse patch features from two frozen pathology foundation models before MIL, with UNI2 as the anchor and CONCHv1.5 as the auxiliary. It first learns an orthogonal alignment, fitted on paired training patches without labels, that maps the auxiliary features into the anchor's feature space. It then splits the anchor features into the part the auxiliary can reach and the remaining complement, averages anchor and auxiliary only inside the reachable part, and keeps the complement unchanged at its original scale, so the fused feature has the same dimension as the anchor. The authors argue this is the smallest possible change to the anchor features that is consistent with the mixed part. Experiments cover BRACS-7, CamelyonPlus, PAM50 and few-shot NSCLC with ABMIL, DSMIL and TransMIL, plus two non-pathology tasks (Stanford Cars and MM-IMDb).

Critical strengths

It's really simple. It's closed-form, needs no labels, and keeps the anchor's dimensionality, so it drops into an existing MIL pipeline without touching the architecture.

The controls are thoughtful: the complement-scale sweep over λ\lambda, Reachable-only, Random-Q orientations, and the adapter study. The authors also report paired bootstrap intervals instead of just bold numbers, which I appreciate.

The writing is careful about what's claimed. "Competitive fixed default" rather than "optimal", and the full-Virchow2 results are openly called mixed.

α=0.5\alpha = 0.5 is fixed a priori, with no tuning on test data.

Critical weaknesses
  1. The gains over the anchor are within noise. Every bootstrap interval against UNI2 contains zero (App. D). The gains (about 0.010.01–0.040.04 BA) are smaller than the std (about ±0.04\pm 0.04–0.050.05), and the BRACS test set has only 31 patients.

  2. It's unclear that the fitted alignment matters.

    • Random-Q is at chance on retrieval, yet it reaches 0.48290.4829 on BRACS ABMIL, only about 0.020.02 below PPFuse.
    • By Eq. 14, PPFuse and full-space interpolation differ only by an invertible linear map, which the first linear layer of the MIL model can absorb.
    • The λ\lambda sweep zig-zags within noise.
  3. Some numbers don't match. Table 15 / Fig. 2(c) report 0.43520.4352 (PPFuse) and 0.36190.3619 (Random-Q) for what looks like the same setting where Tables 2 and 4 report 0.50620.5062 and 0.48290.4829.

  4. The scope is narrow and the baselines are limited.

    • Main results use a single encoder pair.
    • With Virchow2 as the anchor, PPFuse is often worse than raw Virchow2.
    • There's no comparison with existing pathology-model fusion methods (prediction ensembles, Meta-Encoder, AdaFusion, COBRA).
  5. There are no intermediate results such as heatmaps.

    • The paper shows no attention heatmaps or region-level analysis, so we can't tell whether fusion shifts the model's focus toward diagnostically relevant regions.
    • Since the slide-level gains are within noise, this kind of evidence is needed to support the complementary-information claim.
Questions for authors
  1. What explains the gap between Table 15 / Fig. 2(c) (0.43520.4352, 0.36190.3619) and Tables 2 and 4 (0.50620.5062, 0.48290.4829)?
  2. If you put a learnable linear layer in front of MIL, does the difference between λ=1\lambda = 1 and λ=α\lambda = \alpha go away? This would directly test whether "native scale" matters beyond optimization.
  3. Can you show that PPFuse beats UNI2 by a statistically significant margin in any setting, for example with more training seeds or a larger test cohort?
  4. How does PPFuse compare with a simple slide-level ensemble of UNI2 and CONCHv1.5 MIL models?
  5. Is there a principled way to choose the anchor, given the mixed Virchow2 results?
  6. Can you show attention heatmaps for UNI2, CONCHv1.5 and PPFuse features on the same slides (e.g., BRACS and CamelyonPlus)? Where lesion annotations exist, can you add a quantitative check of how well the high-attention regions overlap the annotated lesions? Do the fused features shift attention toward diagnostically relevant regions?
  7. Which patches or tissue types change most after fusion, and do they line up with the regions where PPFuse's predictions differ from UNI2's?
Reviewer_smTYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PPFuse addresses fusion of paired frozen representations with unequal feature dimensions, primarily from two pathology foundation models encoding the same tissue patches. It fits a rectangular semi-orthogonal Procrustes map using training features without task labels. This map decomposes the higher-dimensional anchor into an auxiliary-reachable subspace and its orthogonal complement. The method mixes scale-matched anchor and auxiliary coordinates within the reachable subspace and preserves the anchor complement at its original scale. The authors show that this reconstruction uniquely minimizes the Frobenius distance to the anchor for fixed fused reachable coordinates, and explain its invertible relationship to full-space interpolation when α>0\alpha>0. Experiments evaluate three multiclass WSI tasks with three MIL backbones, few-shot TCGA-NSCLC classification, alternative encoder configurations, and two non-pathology tasks. The paper also studies complement removal and scaling, random versus fitted alignment, and downstream adaptation. PPFuse exceeds both individual encoders in mean balanced accuracy in eight of nine main WSI settings and ranks first among the evaluated methods in seven.

Critical strengths

S1. A simple, mathematically transparent fusion rule with a useful distinction between information preservation and component scaling. The paper clearly separates alignment, reachable-coordinate mixing, and complement treatment. The minimum-change reconstruction is well defined, and the resulting transform retains the anchor dimensionality without training a supervised fusion module. Importantly, Equation 14 explicitly acknowledges that PPFuse and full-space interpolation contain the same linear information for α>0\alpha>0) This avoids conflating a favorable representation scale with additional information or expressive power. The practical contribution is therefore easy to understand and implement: an explicit, label-free choice of representation geometry before supervised downstream learning.

S2. Informative controls and candid reporting of the empirical limitations. The experiments examine more than the final accuracy table. Complement removal, the complement-scale sweep on two datasets, random semi-orthogonal maps, and adapter controls address distinct aspects of the proposed mechanism. The substantial difference between fitted and random maps in patch retrieval, together with the smaller downstream classification gap after adaptation, is particularly informative about the distinction between geometric correspondence and predictive utility. Across three MIL backbones and several tasks, the generally favorable mean results support PPFuse as a potentially useful practical default.

Critical weaknesses

W1. The methodological advance is modest The method combines classical Procrustes alignment with an elementary constrained projection. Its distinguishing choice over full-space interpolation is to set the complement coefficient to λ=1\lambda=1 rather than λ=α\lambda=\alpha. As the authors correctly show in Equation 14, these representations are related by an invertible linear transformation for α>0\alpha>0, so a downstream model with an unconstrained affine input layer can represent the same predictions from either one. This does not invalidate PPFuse, but it makes the empirical behavior under optimization, regularization, and input normalization central to its significance.

W2. The practical benefit of its central scaling choice remains uncertain. The scale sweeps are helpful, yet they use the common downstream training recipe rather than testing whether the preference survives reasonable changes to that recipe. Moreover, all six paired BA intervals comparing PPFuse with full-space OP interpolation on BRACS and PAM50 include zero (Tables 11–12), as do all nine main-task comparisons against UNI2 (Table 7). These intervals do not establish equivalence or the absence of a benefit; they indicate that the magnitude and reliability of the gains remain imprecisely estimated. My reservation is therefore about the strength of the practical evidence for this specific default, rather than the correctness of the reconstruction theorem.

W2. The comparison omits practical multi-encoder fusion approaches. The paper evaluates useful generic baselines, including concatenation, PCA, CCA, and a learned MLP, but does not directly evaluate the pathology fusion approaches discussed in Section 2, such as Meta-Encoder or AdaFusion. The distinction between label-free feature construction and supervised fusion is legitimate, but the eventual application trains a supervised downstream predictor in either case. A comparison with at least one representative learned fusion method, and a simple prediction-averaging ensemble, would help determine the practical accuracy–cost trade-off.

Questions for authors

Please refer to the strengths and weaknesses above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.