acceptodds
Under review as a conference paper at ICLR 2027

MultiMagNet: Multi-Magnification Fusion with Availability Masking, Magnitude Dropout, and Feature Alignment Regularisation for Robust Breast Cancer Histopathology Classification

Abstract

Automated breast cancer diagnosis from haematoxylin and eosin (H&E) histopathology slides is limited by a magnification-scale gap: almost all deep learning systems train at a single, arbitrarily chosen zoom level, discarding the multi-scale information that pathologists routinely combine across , , , and magnification. We introduce MultiMagNet, a four-stream late-fusion framework that jointly encodes all four magnifications through frozen ImageNet-pretrained ResNet-50 backbones augmented with three components: (i) availability masking, an explicit binary signal distinguishing a genuinely absent stream from an informative zero-valued embedding; (ii) magnitude dropout, which stochastically erases entire streams during training (), implicitly ensembling over all missing-view patterns at essentially zero extra cost; and (iii) feature alignment regularisation (FAR), an -alignment auxiliary loss enforcing coherence between co-available stream embeddings. On BreakHis (7,909 images, 82 patients), under strict patient-level partitioning, naive four-stream fusion already exceeds the strongest single-stream baseline (, AUC ) by 4.0 AUC points. Adding our proposed components yields a set of trade-offs rather than a single optimal model: magnitude dropout provides the best robustness to missing views, while a gated variant (E10) reaches the highest full-stream AUC-ROC () but is fragile under partial availability. Transfer to the BACH dataset yields a macro AUC of ; because BACH is single-magnification and the task is four-class, this is not directly comparable to the binary BreakHis result, but demonstrates retained discriminative capacity out-of-distribution. We evaluate the framework's clinical reliability using Monte Carlo Dropout uncertainty decomposition, Mahalanobis out-of-distribution detection (AUROC ), temperature calibration, and stream attribution methods, resulting in a robust suite of engineering choices deployable at an inference cost of ms per sample on a single GPU.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multi-magnification fusion, histopathology image classification, breast cancer diagnosis, missing modality robustness, uncertainty quantification, out-of-distribution detection, attention gating, medical image analysis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Pr6KOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper does not establish a strong motivation for the proposed method. Although pathologists may inspect tissue at multiple magnifications, the paper does not demonstrate that the four magnification streams used in BreakHis represent a realistic multimodal acquisition setting. The images are not spatially registered, and the missing-view scenario is largely simulated by zeroing out streams. Consequently, it remains unclear whether the proposed task reflects an actual clinical bottleneck or is mainly an artificial benchmark constructed from an existing dataset.

Critical strengths

NA

Critical weaknesses

The paper does not introduce a methodological contribution. Availability masking is a standard way to encode missing modalities, magnitude dropout is related to modality dropout or ModDrop, late concatenation and soft-attention gating are well-established fusion mechanisms, and the proposed feature-alignment regularizer is essentially a normalized embedding-distance or cosine-similarity penalty. Combining these components into a single framework may be practically useful, but the paper does not explain why this combination should be considered a research contribution.

Questions for authors

Refer to [Weaknesses]

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

This paper addresses the magnification-scale gap in breast histopathology: most systems train and evaluate at a single, arbitrarily chosen magnification, discarding the multi-scale cues pathologists routinely combine across 40×/100×/200×/400×. It proposes MULTIMAGNET, a four-stream late-fusion model in which each magnification is encoded by a frozen ImageNet ResNet-50 with a small trainable projection head and the four embeddings are fused by masked concatenation or masked softmax gating, augmented by an availability mask, stream-level dropout during training, and an ℓ2 feature alignment loss. On BreakHis under strict patient-level partitioning, naive four-stream fusion improves AUC from 0.890 for the best single stream (200×) to 0.930, and the gated variant reaches 0.935 ± 0.003; a sweep over all 15 non-empty missing-magnification patterns shows the gated model is the most accurate when data are complete but the most fragile when streams are absent, leading the authors to recommend a concatenation variant for uncertain acquisition.

Critical strengths
  • Patient-level partitioning with an explicit assertion check, and an explicit call-out that image-level splitting leaks patient identity and inflates accuracy across most of the BreakHis literature. This alone makes these numbers more trustworthy than the majority of published results on this dataset.
  • Evaluating all 15 non-empty stream subsets, rather than the complete- or single-dropout protocols usual in multimodal work, is more systematic than typical and does surface a real difference between the fusion variants.
  • The derivation that FAR's gradient is orthogonal to the normalised embedding — hence purely rotational, never rescaling, and vanishing as streams align — together with the angular bound of Proposition 1, is the most rigorous theoretical component of the paper and correctly predicts the observed saturating behaviour instead of a collapse to a point.
  • MC-Dropout aleatoric/epistemic decomposition, calibration, Mahalanobis OOD detection, three complementary attribution methods, decision-curve analysis, and a complete parameter/FLOP/latency profile. Coverage of this kind is above average for a submission of this type.
  • Well-organised appendices, fixed seeds, precise metric definitions, and source code in the supplementary material.
Critical weaknesses
  1. No comparison to any published method, in a field where multi-scale fusion is already a mature line. The only non-self model in Table 1 is CSPResNet50, a generic CNN backbone implemented by the authors at a single magnification and — by their own admission — not exhaustively tuned. Every method discussed in Related Work (Spanhol et al. 2016; Bayramoglu et al. 2016; CLAM, DSMIL, HIPT) is absent from the experimental table. Most importantly, pathology foundation models (UNI2, CONCH, Virchow2) are not evaluated even though the authors themselves state in the Limitations that these are the most direct route to higher accuracy. The only external anchor is Spanhol's 2016 figure of 80–85%. This is particularly hard to excuse in computational pathology, where combining scales is not new: hierarchical and multi-resolution MIL, multi-scale encoders and zoom-tree models are an established direction, so the contribution here is a particular instantiation of a known idea rather than the idea itself — which makes comparison to that line more necessary, not less. Under patient-level protocols the published anchor points are sparse — for example 83.95% accuracy at 200× for a patient-wise custom CNN (2025) and 84.3% at 100× for DenseNet+SE with multi-level transfer (2024) — and both are single-magnification results, so neither answers how a multi-magnification fusion system compares. The present model reaches 89.7% accuracy / 0.935 AUC, but because no published method is reproduced under this protocol, the reader cannot locate this number relative to the field.
  2. BreakHis has 82 patients (81 parsed identifiers). The test partition of 1,288 images therefore corresponds to on the order of a dozen patients (inferred from the partition sizes; the exact patient count of the fixed split is not reported). The reported seed-to-seed standard deviations (±0.003 AUC) consequently understate the real uncertainty by more than an order of magnitude. The authors' own cross-validation makes this explicit: fold-mean AUC ranges from 0.833 to 0.995, with s.d. 0.059 across fold means, and because the three seeds reuse the same folds the bootstrap CI is optimistic by the authors' own admission. A patient-level bootstrap is needed throughout. At the subtype level the situation is worse: phyllodes tumour has 3 patients and lobular carcinoma 5, so no subtype-level conclusion is supportable.
  3. Only one dataset; a motivation–data mismatch; and a method tied to exactly four scales that is never tested at any other scale count. BACH is not a second dataset: 80 test images, single magnification, four classes, and the authors state it is not directly comparable. I accept the real constraint — publicly available multi-magnification breast H&E cohorts are scarce — and the authors deserve credit for raising it. But multi-magnification data is not actually scarce, because WSIs are image pyramids: sampling 40/100/200/400× patches at the same coordinate from any WSI yields a second multi-magnification cohort. Notably, BACH Part B itself supplies 40 pixel-wise annotated WSIs, which the authors had in hand and did not use; the same applies to Camelyon16/17 and TCGA-BRCA, and to BRACS (a WSI-derived ROI collection, whose underlying WSIs may need to be requested from the curators). Two further consequences follow. First, Section 1 motivates the work with the shortage of pathologists in sub-Saharan Africa (Wilson et al., Lancet 2018), yet the data are entirely from Brazil (P&D Laboratory, Paraná) and Portugal (Ipatimup, Porto / Castelo Branco): the deployment setting named in the motivation is never evaluated. Second, the framework is instantiated with exactly four streams because BreakHis happens to have four magnifications, yet nothing in the architecture requires four — two streams would work identically, as would six. Because there is only one dataset, the method's dependence on the number of scales is never probed: there is no two-scale or three-scale variant, and the 15-pattern sweep is only ever exercised in the four-scale regime. We therefore learn that the method works on BreakHis, but nothing about whether the fusion gain, the robustness ordering, or the gated-model fragility persist when fewer (or more) scales are available — which is precisely the regime a deployable system would face.
  4. The gating head is the smallest learned component, so the headline negative result may reflect capacity, not gating. The module that scores each stream is a single hidden-layer MLP of width 64 over a 128-d embedding — the weakest learning component in the whole model. Two confounds therefore prevent the most interesting result of the paper from being attributed to the gating mechanism itself: (i) capacity — a 64-wide scorer may simply be unable to learn stable stream preferences, particularly under the distribution shift induced by stream erasure; and (ii) input dimensionality — gating also reduces the classifier input from 516-d to 132-d, so E10 differs from the concatenation models in mechanism and input size simultaneously. The conclusion that gating is brittle under missing views is consequently over-stated relative to what was tested.
  5. Two of the three headline contributions have negligible or negative measured effect: FAR changes AUC from 0.928 to 0.926 (−0.002); its only demonstrable benefit is reduced seed variance (±0.001 AUC, ±0.007 MCC). Availability masking is worth +1.7 points, and only under partial availability. Presenting both as headline contributions alongside a fusion framework is not proportionate to the observed effect sizes.
Questions for authors

Please refer to weaknesses.

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

The paper introduces MultiMagNet, a four-stream late-fusion framework that combines features from frozen ImageNet-pretrained ResNet-50 backbones augmented with three components: availability masking, magnitude dropout, and feature alignment regularisation. Experiments assess classification performance, robustness to missing magnifications, and transfer to another dataset.

Critical strengths

The paper addresses a practically relevant problem about integrating information across magnifications while accommodating missing views. The evaluation over all 15 nonempty view subsets is a useful experimental contribution. The experiments are clearly organized and cover classification, missing-view robustness, calibration, uncertainty, and interpretability. Patient-level partitioning with cross-validation shows some rigor, and the ablation study helps distinguish the effects of individual components.

Critical weaknesses
  1. The methodological novelty appears limited. Magnification effects and multiscale feature integration have been extensively studied in many histopathology classification studies, including through feature fusion and attention-based approaches. From my perspective, the potentially distinctive aspect is how missing magnifications are handled in this work. However, the methodological advance over existing multiscale and missing-view learning approaches remains unclear. The authors should provide a more comprehensive literature review and direct comparisons with relevant methods.

  2. The improvements beyond naive concatenation are modest. In Table 1, E10 improves over E5 from 0.930 to 0.935 in AUC and from 0.892 to 0.897 in accuracy. Other components do not consistently improve performance. For example, adding FAR to E8 reduces AUC from 0.928 to 0.926.

  3. Full-stream prediction requires images at four magnifications and processes them through four encoders. Section 5.7 reports substantially higher computation and inference time than the single-magnification baseline (GFLOPS: 31ms vs. 6.8ms), raising questions about the claimed suitability for resource limited deployment. The authors should establish whether the performance gains justify the additional computational costs.

  4. Relatedly, the choice of four magnifications is insufficiently justified. Although these magnifications are usually available for some datasets and capture different tissue details at different levels, the paper does not establish whether all four are necessary or more magnifications can contribute. Comparisons with models trained on different choices would clarify the performance cost trade off.

  5. Some reported metrics require clarification. Section 5.3 gives E10’s minimum subset accuracy as 0.651 and its spread (max-min) as 0.187. These imply a maximum accuracy of 0.838, below the accuracy of 0.897 reported in Table 1.

  6. Section 3.2, in the paragraph following Equation (6), claims that FAR avoids representation collapse to a point. However, encouraging different magnifications of the same sample to look similar does not guarantee that different samples remain distinguishable. Although the classification loss may help prevent collapse, the claim about FAR itself needs some further clarification.

Questions for authors

Please refer to the questions raised in the Weaknesses section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.