acceptodds
Withdrawn from ICLR 2027

Geometry-aware Multiple Instance Learning via a Shared Riemannian Metric for Few-Shot Whole-Slide Image Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The diagnostic paradigm of whole slide images (WSIs) involves a frozen pathomorphological model, providing patch features, while multiple instance learning (MIL) accomplishes aggregation under weak slide-level labels. Nevertheless, this paradigm has two issues in practice: feature manifold collapse and unstable attention. By analyzing them, we realize that both belong to the two ends of the same causal chain, where the disruption of the feature manifold causes oscillatory changes in the attention distribution that depends on the tangent-space component. Based on this finding, we propose geometry-aware multiple instance learning (GeoMIL), which simultaneously drives feature transformation and attention generation with a single learnable Riemannian metric. We design a tangent space low-rank projection (TSLP) module to perform task adaptation while maintaining feature geometry, and use geometric query attention (GQA) to anchor the attention distribution to the tangent space stabilized by TSLP. Through extensive experiments, we demonstrate that our method maintains superior few-shot WSI classification performance while achieving a more stable attention distribution.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-instance learning, whole-slide images, few-shot learning, computational pathology

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

This paper proposes GeoMIL, a geometry-aware multiple instance learning framework for few-shot whole-slide image classification. It addresses feature manifold degradation and attention instability through a shared learnable Riemannian metric. Specifically, TSLP preserves feature geometry while enabling task adaptation, and GQA generates attention queries from geometric representations. Experiments on three pathology datasets demonstrate improved classification performance and attention-based localization.

Critical strengths
  1. The paper identifies a meaningful connection between feature manifold degradation and attention instability in few-shot WSI classification, providing a geometry-based perspective on existing MIL limitations.
  2. The paper provides comparative experiments against ten representative MIL methods on three pathology datasets, together with ablation studies and attention localization analyses to demonstrate the effectiveness of the proposed framework.
Critical weaknesses
  1. The claimed causal relationship between feature manifold degradation and attention instability is not sufficiently supported. Specifically, Figure 1(c) measures attention concentration rather than instability across epochs. A more appropriate metric would be the Jensen–Shannon divergence between consecutive attention distributions, Δt=DJSD(αt,αt−1)Δ_t = D_{JSD}(α_t, α_{t−1}). The attention concentration curves in Figure 1(c) rapidly increase and then stabilize, without clearly demonstrating the claimed oscillations.
  2. The computational efficiency of GeoMIL is insufficiently evaluated. The proposed framework introduces additional Riemannian metric pretraining and a teacher–student architecture, but the paper provides no comparisons of training time, inference time, or computational complexity with existing MIL methods. Such analyses are necessary to assess the practical overhead of the proposed approach.
  3. A decreasing effective rank does not necessarily indicate harmful manifold degradation, as it may reflect beneficial discriminative feature compression that suppresses task-irrelevant variations.
  4. The practical validity of the assumptions underlying Theorem 1 remains insufficiently examined. Moreover, the theoretical analysis only establishes the geometric preservation properties of the feature transformation, without directly demonstrating that preserving feature geometry leads to improved attention stability.
Questions for authors
  1. Further experiments directly measuring attention fluctuations and establishing the causal role of feature geometry are needed to substantiate the central motivation.
  2. Please clarify which evaluation metric is used as the criterion for early stopping.
  3. The independent contribution of GQA appears limited. Although GQA generally improves performance, some gains are relatively small compared with the reported standard deviations across random seeds. Paired statistical significance tests and direct comparisons of attention fluctuations with and without GQA are needed to better establish its effectiveness.
  4. Please provide additional experimental results on multi-class classification datasets to further demonstrate the generalizability of GeoMIL.
Reviewer_ZzMsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoMIL first pretrains a metric network on unlabeled support-set patches with a Riemannian metric-matching objective. For each patch, the network outputs a low-rank matrix meant to approximate a projection onto the local tangent space of the feature manifold. GeoMIL uses this network in two ways inside ABMIL. In the tangent-space linear projection (TSLP), patch features are projected with a frozen copy of the metric network (the teacher) plus a small trainable correction (the student), and an extra tangent loss keeps that correction small. In the geometry-aware query (GQA), the attention query is computed from the average of the projected patch features, so each slide gets its own query. Experiments cover CAMELYON-16, TCGA-NSCLC and TCGA-BRCA in the 2/4/8/16-shot settings, with CONCH features and five seeds.

Critical strengths
  1. The motivation is interesting. Few-shot MIL heads overfitting through a geometry-agnostic linear layer is a real concern, and tying the feature projection and the query to one learned geometric object is a neat design idea.
  2. The gains at very low shot on the sparse-positive CAMELYON-16 are fairly large (4-shot AUC 62.862.8 vs. 58.058.0 for FOCUS; 8-shot 79.879.8 vs. 75.675.6).
  3. The projection-source ablation (Metric vs. Random / PCA / untrained Linear) and the query ablation are the right controls, and the paper reports sensitivity to n0n_0 and rr.
Critical weaknesses
  1. The main claim isn't tested. The "causal chain / more stable attention" claim rests only on a correlational analysis of baseline ABMIL (Fig. 1). Attention stability is never measured for GeoMIL itself, and the localization gains are under 2 Dice.

  2. There's a conceptual gap in the geometry. The tangent projector is applied to the position vector rather than to a displacement, so the result depends on where the coordinate origin is and isn't geometrically meaningful. App. B.3 also contradicts TSLP's motivation about staining and scanning factors.

  3. The baselines look under-tuned and incomplete. ASMIL is below plain ABMIL in every cell, and MAMMOTH, TOP and FAST aren't compared. The claimed "consistent lead" also fails at 16-shot NSCLC.

  4. Hyperparameters are chosen on test, and the cost is unreported. rr and n0n_0 are selected from test-split sweeps, and the compute cost of the per-patch metric network isn't given. Removing the tangent loss even improves 4-shot CAMELYON.

Questions for authors
  1. Can you measure attention stability (e.g., epoch-to-epoch DJSDD_\text{JSD}, or attention variance across seeds) for GeoMIL vs. ABMIL / ASMIL / MR-ABMIL? That's the paper's main claim.
  2. Why project xx rather than a displacement like x−xˉx - \bar{x}? Does using M(x)(x−xˉ)M(x)(x - \bar{x}) change the results?
  3. How were the baselines tuned, and why is ASMIL below ABMIL everywhere?
  4. How does GeoMIL compare with MAMMOTH and with TOP / FAST under the same splits?
  5. What are the training and inference cost and memory of the per-patch metric network? What n0n_0 do you use at 16-shot?
Reviewer_pvRaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GeoMIL, a multiple instance learning framework for few-shot whole-slide image classification. The authors argue that feature manifold collapse and attention oscillation across epochs are linked through the distortion of the feature manifold tangent space by unconstrained linear layers. GeoMIL pretrains a low-rank Riemannian metric factor network in Phase 0 using conditional carre du champ matching on unlabeled patch features from the support set. In Phase 1, Tangent-Space Low-Rank Projection freezes a teacher metric network and trains a student network regularized by a projection consistency loss, while Geometry-Query Attention computes the attention query vector from the slide-level average of the teacher projections. The method is evaluated on CAMELYON-16, TCGA-NSCLC, and TCGA-BRCA under 2, 4, 8, and 16-shot settings using frozen CONCH features.

Critical strengths

The paper investigates an important practical failure mode in few-shot multiple instance learning, namely the rapid overfitting of linear projection layers and the resulting concentration of attention weights on background or non-diagnostic tissue regions. Connecting self-supervised Riemannian metric estimation to regularized adapter projections is an interesting conceptual direction.

Critical weaknesses
  1. The manuscript relies heavily on LLM-generated initial drafting, resulting in repetitive, verbose prose and poor structural readability. As acknowledged in the AI Use Statement on page 10, generative AI was used for the preparation of the initial draft. This is evident throughout the text: the writing repeats grand, formulaic slogans such as two ends of the same causal chain across the abstract, introduction, method, and conclusion without building clear, rigorous transitions. More importantly, the presentation of the core diagnostic experiment in Section 1 and Figure 1 is severely disconnected from its actual mathematical definition in Appendix B. In the main text, Figure 1 is presented using undefined notations and vague terms like tangential energy proportion, forcing the reader to flip back and forth to pages 15 and 16 just to understand what features are being projected and how the curves are computed.

  2. The core mathematical formulation in Section 3.3 and Theorem 1 conflates local tangent vectors with ambient position coordinates, and omits how the noise scale conditioning is handled during training and inference. According to Property 1 in Appendix A, as the noise scale approaches zero, the product of the low-rank metric factor with its transpose converges to the orthogonal projection operator onto the local tangent space of the manifold at a given point. In differential geometry, the tangent space at a point is a linear subspace of local displacements attached to that point, not a subspace of absolute ambient coordinate vectors. Yet Equations 4, 5, and 6 as well as part iv of Theorem 1 apply the local differential operator directly to the raw position vector itself, claiming that decomposing the raw position vector into tangent and normal components removes normal-space artifacts and preserves tangential semantics. For a nonlinear manifold embedded in ambient space, the raw coordinate vector is not an element of the local tangent space at all; for instance, on a sphere centered at the origin, the position vector is strictly orthogonal to the tangent space everywhere, so its tangential projection is identically zero, and projecting raw position vectors depends entirely on the arbitrary choice of the coordinate origin. Furthermore, Section 3.2 states that the metric network takes the noise bandwidth as a conditioning input via FiLM modulation, yet the paper never specifies the sampling distribution of this bandwidth during Phase 0 pretraining nor what numerical value is fed into the teacher and student networks during Phase 1 when evaluating on clean, unnoised features.

  3. The diagnostic motivation in Figure 1 and Appendix B directly contradicts the proposed method, and the Phase 0 scaling ablation in Table 6 refutes the manifold learning claim. In Figure 1, panels a and b show that on a normal slide, standard attention concentration coincides closely with patches having high tangential energy, and panel c shows that this erroneous attention concentration is driven by the tangentially projected features rather than the normal component. Crucially, as specified in Appendix B.2 on pages 15 and 16, the local PCA tangent projection in Equations B1 and B2 is computed directly on the frozen foundation model input features before any linear layer is applied. Thus, Figure 1 actually demonstrates that the frozen foundation model's own local tangent space is already dominated by high-variance non-diagnostic variation on normal slides. However, Phase 0 of GeoMIL is trained strictly on these exact same unlabeled frozen features to preserve their tangent space and annihilate their normal space. This fundamental contradiction is confirmed in Table 6 on page 18, where increasing the number of unlabeled slides in Phase 0 from 4 or 8 to 16 or 32 consistently degrades classification accuracy across datasets, such as 8-shot TCGA-BRCA AUC dropping from 93.6 to 90.3 and F1 dropping from 74.5 to 71.7. If Phase 0 pretraining were genuinely estimating the intrinsic semantic manifold of pathology features, providing more unlabeled slides from the same domain would improve the tangent space estimate rather than harming performance. In addition, Appendix C.3 reveals that while only 2 to 16 slides per class are used for training, 20 percent of all remaining slides in the dataset, amounting to roughly 80 to 220 labeled slides, are used as a validation set for early stopping, which violates the few-shot evaluation premise.

Questions for authors
  1. Clarification request on the geometric formulation and noise conditioning: Since the learned metric operator projects local displacements rather than absolute ambient coordinates, how do the authors justify multiplying the metric factor directly by the raw feature vector in Equations 4 to 6? What sampling distribution of the noise bandwidth is used in Phase 0, and what exact bandwidth value is passed to the FiLM conditioning layer of the teacher and student networks during Phase 1?

  2. Clarification request on Figure 1 and Appendix B.2: In Equations B1 and B2, the tangent projection is estimated on the frozen input features, which do not change across epochs. Since Figure 1 shows that the tangent space of the frozen input features is precisely what drives false attention concentration on normal slides, why would training Phase 0 on those same frozen input features to preserve their tangent space suppress non-diagnostic attention?

  3. Clarification request on the few-shot validation split: How do GeoMIL and the baselines perform when early stopping is conducted strictly on a validation split subsampled from the few-shot support set, or without early stopping at a fixed epoch, rather than using 20 percent of the entire remaining dataset for checkpoint selection?

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

Related papers

Discussion (0)

Sign in to comment.