acceptodds
Withdrawn from ICLR 2027

GeoGCD: Geometry-Aware Prototype Learning for Histopathology Generalized Cancer Category Discovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generalized Category Discovery (GCD) aims to simultaneously recognize labeled categories and discover novel categories from partially annotated data. Histopathology generalized category discovery is particularly challenging due to the subtle morphological differences among cancer subtypes and the substantial appearance variations within the same subtype, which often lead to ambiguous decision boundaries and less discriminative feature representations. Existing GCD methods mainly focus on classifier optimization or clustering objectives while paying limited attention to preserving fine-grained histopathological structures and exploiting the geometric relationships among category prototypes.To address these challenges, we propose GeoGCD, a geometry-aware prototype learning framework for histopathology generalized category discovery. Specifically, we introduce a Feature Detail Refinement and Aggregation (FDRA) module that enhances discriminative local structural representations by preserving fine-grained pathological details while suppressing redundant feature responses. Furthermore, we develop an Adaptive Margin Prototype Adapter (AMPA) module that dynamically calibrates classification margins according to the geometric relationships among category prototypes, encouraging better discrimination between confusing cancer subtypes. By jointly improving local structural representations and prototype distributions, GeoGCD learns more compact intra-class representations and better-separated inter-class prototypes, leading to more reliable novel category discovery.Extensive experiments on multiple histopathology benchmark datasets demonstrate that GeoGCD consistently outperforms existing generalized category discovery methods under realistic semi-supervised settings, validating its effectiveness for discovering previously unseen cancer categories.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generalized Category Discovery, Medical Image Analysis, Semantic-guided Learning, Fine-grained Representation.

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_HAX7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes GeoGCD, a method for Generalized Category Discovery (GCD) in histopathological images. Building upon the semi-supervised learning framework of SimGCD, the authors introduce two modules: FDRA and AMPA. FDRA generates channel-gating mechanisms from the high-frequency components of features and integrates fine-grained details with global information via cross-attention. AMPA establishes adaptive classification margins for labeled samples by calculating the cosine similarity between each class prototype and its nearest neighboring prototype. Experiments conducted on the LC25000, BreakHis, and BACH datasets demonstrate that GeoGCD outperforms SimGCD.

Critical strengths
  1. The use of high-frequency information in FDRA to modulate raw features is an intriguing concept.

  2. AMPA’s approach, determining margins based on the nearest competing class, is a logical strategy for addressing local inter-class confusion.

  3. According to the reported metrics, the full model shows improvements in overall, known-class, and novel-class accuracy, suggesting that the combined modules may enhance the balance between these two objectives.

Critical weaknesses
  1. The criteria for matching (used in evaluation) are not clearly defined. In Table 1, the results for GeoGCD on the LC25000 dataset are: All=72.6, Seen=76.1, and Novel=97.7; for BreakHis, they are 67.2, 81.9, and 69.0, respectively. In both cases, the "All" (overall) accuracy is lower than the accuracy for each individual subset. The paper mentions the use of Hungarian matching but fails to explain how the matching process is handled across these three metrics. If these metrics are calculated based on the same evaluation set and a single global matching assignment, the overall accuracy should not be lower than the accuracy of both subsets. The authors should provide the grouped results and confusion matrices derived from a consistent global matching.

  2. The paper lacks sufficient detail and often reads like a mere list of formulas, making the overall line of reasoning unclear. Furthermore, there is no supplementary material to clarify points left ambiguous in the main text (which, while not strictly mandatory, is problematic given that the main text itself is not sufficiently detailed). The quality of the figures is also mediocre.

  3. The current ablation studies only demonstrate that performance changes when the modules are added, they do not establish whether these two core designs are truly indispensable. The paper lacks further analysis to address this gap. Figure 6 reports overall intra-class and inter-class distances, however, it lacks a precise definition of these distances, fails to specify the datasets used, and does not provide a breakdown for "known-known," "known-new," and "new-new" class pairs. Consequently, it is impossible to determine which specific aspects of the geometric structure have improved. Furthermore, the overall performance gains are marginal: the complete model shows improvements of only 0.3, 0.8, and 0.3 percentage points over the baseline. Meanwhile, the addition of FDRA alone causes the overall accuracy to drop from 72.3% to 71.6% on the LC dataset, and from 73.8% to 72.1% on the BACH dataset. While the authors highlight improvements in metrics for new classes, they fail to adequately address the implications of the decline in overall performance.

  4. The experimental details are also insufficient. The experimental section merely states that a standard GCD (Generalized Category Discovery) protocol is followed, but it fails to systematically specify the lists of known and new classes for each dataset, the labeling ratios, the sample sizes for each subset, or the methodology for repeating class splits. Moreover, it remains unclear whether the evaluation is conducted on unlabeled data already encountered during training, or on entirely independent samples or patients.

Questions for authors

See "Critical Weaknesses".

And please check the format of citation, there may be something wrong.

Reviewer_dy65OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GeoGCD, a generalized category discovery framework for histopathology images. Building on SimGCD, the method introduces two components: Feature Detail Refinement and Aggregation (FDRA), which combines frequency-domain feature processing, channel gating, and cross-attention; and Adaptive Margin Prototype Adapter (AMPA), which adjusts classification margins based on cosine similarities between category prototypes.

The method is evaluated on three histopathology datasets, LC25000, BreakHis, and BACH, using overall, seen-category, and novel-category clustering accuracy. The authors additionally present component ablations, t-SNE visualizations, Grad-CAM examples, and feature-distance analyses. The reported improvements over SimGCD are relatively small.

Critical strengths

The paper addresses a potentially useful semi-supervised learning setting and provides a modular extension of SimGCD. The two proposed components are evaluated individually and jointly on three public datasets.

Critical weaknesses
  1. Limited methodological novelty and insufficient justification of the histopathology-specific contribution.

The proposed method largely combines established techniques within the existing SimGCD framework. FDRA employs frequency-domain feature processing, channel-wise gating, and cross-attention, while AMPA adjusts classification margins based on prototype cosine similarity. These components are conceptually close to existing frequency-attention and margin-based representation-learning approaches.

Although combining such mechanisms may be practically useful, the paper does not establish a sufficiently distinctive technical contribution. More importantly, it does not demonstrate why this particular combination is necessary for histopathological generalized category discovery.

The proposed frequency-aware refinement is motivated by fine-grained pathological structures, including nuclei and gland boundaries. However, the experiments do not quantitatively establish that FDRA preserves or enhances these structures. The improvement of embedding compactness alone cannot substantiate this claim.

The connection between AMPA and novel-category discovery also requires stronger justification. According to the formulation, the adaptive margin modifies the supervised target-class logit for labeled samples, while unlabeled samples follow the original pseudo-label learning objective. The mechanism by which this design improves novel-category separation is therefore indirect and insufficiently analyzed.

A related concern is the application setting. The paper repeatedly motivates the work through the discovery of previously unseen cancer categories, yet the evaluation consists of artificially withholding labels from predefined classes in existing classification datasets. The authors do not establish whether the intended application involves discovering previously unrecognized pathological subtypes, identifying rare existing diagnoses, or reducing annotation requirements.

These are substantially different tasks. Without a clearly defined application and an evaluation demonstrating its relevance, the claimed histopathology-specific contribution remains weak.

  1. Marginal improvements and substantial concerns regarding evaluation reliability.

The quantitative results provide limited evidence of meaningful improvement over the main baseline. Compared with SimGCD, GeoGCD increases overall clustering accuracy by only 0.3, 0.8, and 0.3 percentage points on LC25000, BreakHis, and BACH, respectively. No repeated-run standard deviations, confidence intervals, or statistical significance tests are reported.

The ablation results also raise concerns about the individual contributions of the proposed components. FDRA alone reduces overall accuracy from 72.3% to 71.6% on LC25000 and from 73.8% to 72.1% on BACH. AMPA alone provides similarly limited improvements. Although their combination performs better, the results do not establish a robust or generalizable benefit.

More importantly, the evaluation metrics require clarification. Under the standard SimGCD evaluation protocol, a single Hungarian assignment is computed over all samples, and overall accuracy is the sample-weighted average of seen- and novel-category accuracies. However, Table 1 reports values such as All = 72.6%, Seen = 76.1%, and Novel = 97.7% for GeoGCD on LC25000. These values cannot arise from the same evaluation population under a shared assignment. Similar discrepancies appear for BreakHis and several baselines.

Different subset-specific assignments or evaluation populations might explain the discrepancy, but the manuscript does not specify such a protocol. In addition, the SimGCD novel-category accuracy on BreakHis differs between Tables 1 and 2.

Dataset splitting is another major concern. LC25000 contains many augmented versions of a much smaller set of original tissue images. Previous research has documented substantial leakage risks when augmented images from the same original tile are distributed across training and evaluation splits. The manuscript does not state whether original-tile-level grouping was enforced. Likewise, BreakHis contains multiple images and magnifications from the same patients, but patient-level separation is not explicitly described.

Finally, the qualitative evidence does not resolve these issues. Although the t-SNE plots and feature-distance analyses show improved embedding separation, this is closely related to the representation-learning objectives themselves. Such results do not establish improved downstream generalization or biologically meaningful category discovery. The Grad-CAM examples are also qualitative, with insufficient explanation of the displayed spatial ground-truth annotations.

Collectively, these limitations substantially weaken the empirical support for the paper's conclusions.

Questions for authors
  1. Could the authors clarify the exact clustering accuracy protocol?

Are the All, Seen, and Novel accuracies computed using a single Hungarian assignment over the same evaluation population, following the standard SimGCD protocol? If so, how can All accuracy be lower than both Seen and Novel accuracy in several rows of Table 1?

If separate assignments or different evaluation populations are used, please provide the precise definitions and justify their comparability with the reported baselines.

Please also explain the discrepancy in the BreakHis novel-category accuracy for SimGCD between Tables 1 and 2.

  1. How were the data split to prevent leakage?

For LC25000, were all augmented images originating from the same original tile assigned to the same partition? For BreakHis, were patients kept disjoint across training, validation, and evaluation, including across magnification levels?

Please report the exact dataset partitions, category assignments, numbers of labeled and unlabeled samples, random seeds, and repeated-run statistics.

  1. What evidence demonstrates that FDRA captures diagnostically meaningful fine-grained structures?

The current evidence primarily consists of Grad-CAM examples and feature-space visualization. Could the authors provide controlled frequency-band experiments and quantitative evidence linking FDRA to relevant histological morphology?

Please also clarify how the spatial feature map used by FDRA is constructed from the DINO ViT backbone, specify the frequency masks, and explain the additional Encoder operation in Eq. 6.

  1. How does AMPA directly benefit novel categories, and how is its adaptive margin controlled?

The current formulation applies the target-class margin to labeled samples. Could the authors analyze the resulting margin distributions and prototype changes separately for seen and novel categories?

Furthermore, the prototype cosine similarity in Eq. 13 can be negative. Is the adaptive margin constrained to be nonnegative? How is the margin coefficient selected?

  1. What practical evidence supports the claimed cancer category discovery application?

All reported novel categories are derived from established labels that are artificially hidden during training. Could the authors clarify the intended histopathology use case and provide evidence that the resulting clusters correspond to meaningful pathological subtypes or can be reliably characterized independently?

It would also be important to evaluate settings where the number of novel categories is unknown, or to test the discovered representations on an independent clinical or biological endpoint.

Reviewer_qM3uOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GeoGCD for generalized category discovery (GCD) in histopathology. It builds on SimGCD and adds two modules. The first, FDRA, applies an FFT high-pass decomposition to the features, reweights them with a channel gate, and fuses the result with the global feature via cross-attention. The second, AMPA, computes each class’s margin from the cosine similarity between its prototype and its nearest neighboring prototype, and subtracts this margin from the target-class logit of labeled samples. Experiments use DINO ViT-B/16 and report All/Seen/Novel accuracy on LC25000, BreakHis, and BACH, together with ablations, Grad-CAM, t-SNE, and a hyperparameter sensitivity analysis.

Critical strengths
  1. GCD for histopathology has practical clinical relevance, and the motivation (high inter-class similarity, large intra-class variation) is reasonable.
  2. The method is a simple plug-in on top of SimGCD and leaves the original training pipeline unchanged.
Critical weaknesses
  1. The reported numbers are internally inconsistent. Under the standard protocol used by SimGCD, All is a weighted average of Seen and Novel and must lie between them. Yet several rows in Table 1 report All below both, e.g., Ours on LC (72.6/76.1/97.7) and on BreakHis (67.2/81.9/69.0), as well as SimGCD, DCR, and ProtoGCD on LC. The authors need to state which protocol they use. If Seen and Novel are computed with separate Hungarian matchings, the numbers are not comparable to those reported by SimGCD and the other baselines.
  2. AMPA has limited novelty, and its scope is questionable. AMPA is essentially a CosFace-style additive margin scaled by the similarity to the nearest prototype. Adaptive margins based on inter-class similarity have been studied before, e.g., the adaptive margin loss for few-shot learning and various class-adaptive margin methods, none of which are discussed.
  3. The gains are small and no variance is reported. Relative to SimGCD, All improves by only 0.3, 0.8, and 0.3 points, whereas GCD results typically vary by 1–2 points across seeds. No mean and standard deviation over multiple runs are reported. In the ablation (Table 2), FDRA alone actually lowers All on LC (72.3→71.6) and BACH (73.8→72.1), which the text does not discuss.
  4. Outdated backbone and baselines: A DINO ViT-B pretrained on ImageNet is an outdated backbone for histopathology; the method should be validated on pathology foundation models such as UNI, CONCH, or CTransPath. Strong recent GCD methods are also missing, e.g., SPTNet, InfoSieve, CMS, μGCD, and LegoGCD.
Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.