acceptodds
Withdrawn from ICLR 2027

RLT: Learning from Relational Long Tails in Whole Slide Image Analysis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-tailed whole slide image (WSI) classification is usually studied from the imbalance of slide labels or sparse instance-level evidence. We argue that this view is incomplete. Diagnostic evidence in pathology is also encoded by interactions among heterogeneous tissues, while these tissue relations are themselves highly imbalanced. Common relations dominate graph propagation, whereas rare but class-discriminative relations can be suppressed, especially for tail classes with few training slides. We therefore propose RLT, a relational long-tail framework for WSI classification. First, we construct a morphology-aware heterogeneous graph in which patch nodes are associated with tissue semantics and relational edges encode tissue interactions. Second, we introduce diagnosticity-aware relation debiasing, which combines relation rarity with counterfactual contribution to enhance rare relations only when they provide class-specific evidence. Third, we develop prototype-guided representation enrichment to densify the sparsely sampled representation regions of data-scarce classes and stabilize the decision space. A class-prior-aware Balanced Softmax objective is further used to reduce residual classifier bias. Experiments on Camelyon+-LT and PANDA-LT demonstrate that RLT consistently outperforms representative baselines in both overall and tail-class performance. Code and configurations are available at: https://anonymous.4open.science/r/RLT_ICLR2027-anonymous/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
long-tailed learning, heterogeneous graph learning, representation learning, whole slide image analysis, computational pathology, prototype learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vJ9TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RLT studies long-tailed WSI classification from a relation-level perspective. It builds a heterogeneous graph using tissue types, estimates the rarity and class-specific diagnostic value of each tissue relation, and uses these scores to adjust graph message passing. It further uses class prototypes and Balanced Softmax to improve tail-class recognition.

Critical strengths
  1. The relation-level view is meaningful. Previous WSI graph methods already model interactions between patches, while long-tail methods mainly focus on class imbalance or rare discriminative instances. RLT connects these two directions and asks whether relation frequency itself creates bias during graph learning.

  2. The method does not simply treat every rare relation as important. It combines relation rarity with counterfactual diagnostic contribution. This is more reasonable than directly increasing the weight of all rare relations.

  3. Experiment is relatively complete. The paper reports five independent runs, component ablations, different imbalance ratios, parameter sensitivity, and tail-class metrics. The gains in Tail-F1 are large on both benchmarks.

Critical weaknesses
  1. It is not fully shown that relation imbalance is an independent source of long-tail failure. The authors say that relation frequencies are highly imbalanced, but this imbalance may partly come from tissue abundance, class frequency, and the way the kNN graph is constructed. A skewed relation distribution alone does not prove that it is a separate cause of poor tail-class performance.

  2. The graph relations are not true spatial tissue interactions. Edges are constructed using Euclidean distance in the feature space rather than physical WSI coordinates. Therefore, a “tumor–stroma” relation may connect two patches that are spatially far apart. The method captures feature-based tissue relations, but terms such as tissue interaction or interface should be used carefully.

  3. The main tail-class improvement is not clearly driven by relation debiasing alone. Diagnosticity-aware relation debiasing by itself gives very limited Tail-F1 improvement. On Camelyon+-LT, Tail-F1 remains 0, and on PANDA-LT it is only 4.2. Adding Balanced Softmax increases these values to 30.4 and 48.5, while prototype enrichment further raises them to 58.15 and 55.36. Therefore, much of the final gain comes from existing class-balancing and feature-augmentation techniques rather than the proposed relational long-tail mechanism itself.

  4. The frozen judge may inherit the same long-tail bias. Relation diagnosticity and inference-time relation gates depend on predictions from a judge model trained on the same imbalanced data. If the judge is already biased toward head classes, this bias may also affect the relation scores. The paper does not provide enough analysis of the judge’s tail-class accuracy or calibration.

  5. The meaning of “prototype enrichment” is somewhat overstated. The method interpolates each representation toward its class mean. This creates additional training points, but it mainly pulls features toward the class center rather than expanding the class manifold. For tail classes with very few samples, the estimated prototype may also be unstable.

Questions for authors

see weakness

Reviewer_3NyMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RLT targets long-tailed slide-level classification. CONCH first assigns each patch one of five tissue types zero-shot (tumor, stroma, immune, necrosis or other). Each patch is then connected to its 4 nearest neighbors in feature space, and each edge is labeled by the pair of tissue types it joins, giving 15 relation types. Each relation type is scored by how rare it is and by a counterfactual contribution, i.e., how much removing it lowers a frozen judge model's confidence in each class. The two are combined into one score per relation type and class; at inference, these class-wise scores are averaged using the judge's predicted class probabilities. The scores are used to amplify the messages along each edge, on top of a residual path. Training also adds a Balanced Softmax loss and a "prototype-guided enrichment" step that randomly pulls a slide embedding part of the way toward the mean embedding of its own class. Experiments use Camelyon+-LT (imbalance ratio about 30, 16 test slides per class) and PANDA-LT (imbalance ratio about 50, 200 test slides per class), with 5 runs on one fixed split, against 11 MIL baselines including the long-tailed methods MDE-MIL and DeCo-MIL.

Critical strengths
  1. The problem matters. Long-tailed slide-level diagnosis (e.g., ITC in Camelyon, high ISUP grades) is clinically relevant and under-studied compared with balanced MIL benchmarks.
  2. Measuring a relation type's class-conditional contribution with a counterfactual, instead of using frequency alone, is a reasonable idea. The paper shows that frequency-only reweighting hurts overall performance (Table 3).
  • The experimental scaffolding is fairly complete:
    • two LT benchmarks with balanced validation and test sets;
    • comparison against two recent pathology-specific LT methods;
    • ablations of graph type, debiasing, prototype design and gating form (Table 7);
    • an imbalance-ratio sweep on PANDA-LT;
    • public code.
  • On PANDA-LT (1200 test slides), the gain in Tail-F1 over the strongest baselines is fairly large: 55.3655.36 vs. 45.3045.30 (ILRA-MIL) and 44.8044.80 (DeCo-MIL).
Critical weaknesses
  1. The relational component isn't what drives the gains. In Table 3, diagnosticity-aware debiasing leaves Tail-F1 at 0.000.00 (Camelyon) and 4.204.20 (PANDA), and almost all of the improvement comes from Balanced Softmax and the prototype term. Frequency-only debiasing even gets a higher Tail-F1 (28.6028.60 / 26.8026.80), and a Graph+BS ablation is missing.

  2. The baseline comparison is unfair. Most baselines appear to use plain CE (Tail-F1 0.000.00 on Camelyon). Standard long-tail controls such as AB-MIL + Balanced Softmax or logit adjustment are missing.

  3. The tail metrics are unreliable. Camelyon has only 16 ITC test slides, and the Tail-F1 std reaches 1818–2828. Several cells are exactly 0.00±0.000.00 \pm 0.00, which suggests the model never predicts the tail class. There's also a single split and no confidence intervals.

  4. The protocol is under-specified. Tail-F1 is never defined. The paper doesn't say whether the single-seed hyperparameter sweep used the validation or the test split.

  5. The "tissue interaction" motivation doesn't match the graph. Edges come from feature-space kNN rather than spatial adjacency, so rare relations may mostly reflect tissue-label noise, and the accuracy of the zero-shot tissue labels is never validated.

  6. There are concerns about the judge and the prototypes. The frozen judge's predictions are biased toward head classes, yet they are used to weight the relation scores at inference. Pulling embeddings toward their class mean shrinks within-class spread rather than "enriching" the tail, and it isn't compared with center loss or mixup.

Questions for authors
  1. Can you add Graph+CE+BS, Freq.Debias+BS and Graph+BS+Prototype to Table 3? How much does diagnosticity-aware debiasing add once BS is present?
  2. Can you report AB-MIL / DS-MIL / DeCo-MIL with Balanced Softmax or logit adjustment under the same protocol?
  3. How is Tail-F1 defined on each dataset? Can you report per-class recall and bootstrap CIs, especially for the 16 ITC test slides?
  4. Were the Fig. 3(c) sensitivity runs and the final hyperparameter choices made on the validation split or the test split?
  5. How accurate are the zero-shot CONCH tissue labels? How do the results change with a spatial-adjacency graph instead of feature-space kNN?
  6. Why does Diag.Debias give 0.000.00 Tail-F1 on Camelyon while Freq.Debias gives 28.6028.60? Why is the Heterogeneous Graph worse than the Homogeneous one on Tail-F1?
  7. Does calibrating the judge (e.g., training it with BS too) change Sr(G)S_r(G) and the final results?
Reviewer_ptVjOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework for long-tailed whole-slide image (WSI) classification that explicitly considers imbalanced tissue relations, called RLT. The method constructs a graph according to patch embedding similarity and categorizes the edges in the graph into 15 relation types according to the tissue categories of the nodes the edges connect. It then introduces diagnosticity-aware relation debiasing, combining relation rarity with counterfactual contributions to reweight messages during propagation. Prototype-guided representation enrichment and Balanced Softmax are then incorporated in the model. Experiments were conducted on Camelyon+-LT and PANDA-LT.

Critical strengths
  1. The paper attempts to utilize the potential role of relationships between different morphological components when analyzing WSI.

  2. Experiments were conducted on pathology benchmarks.

  3. The paper includes ablation studies that investigates the contributions of different components.

Critical weaknesses
  1. The rationale and hypothesis behind the graph construction and message-passing mechanism are unclear. The proposed method constructs a graph according Euclidean distances in the CONCH embedding space. That means a Tumor–Immune edge merely indicates that a tumor-labeled patch and an immune-labeled patch are connected because of visual/embedding similarity, despite of their spatial adjacency. The visualization in Figure 2(a) is simply misleading, as patches from long distances will be connected because of their embedding similarities. There are no biological mechanism that make such two patches interact with each other simply because they look similar. Therefore, it is unclear how the constructed graph is related to biological tissue interactions, and why the message propagation in such a network is expected to be biologically meaningful and to be useful in improve WSI classification.

  2. The central motivation of the paper is that tissue relations exhibit highly imbalanced frequency distributions, with rare relations potentially carrying important diagnostic information.However, these relation frequencies are measured on graphs constructed using feature-space similarity rather than their spatial adjacency. The patches belonging to the same category are more likely to look similar and are more likely to be connected. Heterogeneous relations are relatively rare because the corresponding tissue categories occupy different regions in the embedding space. The reported relational long tail reflects properties of the chosen graph-construction procedure rather than pathological tissue interactions, and is unlikely to imply any diagnostic information.

  3. The proposed enrichment interpolates each graph embedding toward its class prototype. The operation moves representations toward the class center and does not necessarily expand the diversity of the representation space, as claimed in the paper.

  4. The features that aggregated from the graph are pooled somehow and are eventually to be used in the prediction. However, it is unclear how the pooling is achieved.

Questions for authors
  1. What is the rationale of constructing graph according to embedding similarity? What is the biological meaning of aggregating information from similar patches through message passing?

  2. How can connections between feature-similar patches from different tissue categories be interpreted as meaningful biological tissue interactions, particularly when the patches may be spatially distant?

  3. What advantages does the proposed graph construction provide over conventional MIL aggregation or spatially based graph-construction strategies?

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

Related papers

Discussion (0)

Sign in to comment.