acceptodds
Under review as a conference paper at ICLR 2027

ABD-TDA:Attention-Based Deep and Topological Feature Fusion for Histopathological Cancer Classification

Abstract

Accurate histopathological image classification remains challenging due to substantial morphological and structural variability across cancer types. While Vision Transformers (ViTs) capture global visual information through self-attention, they do not explicitly represent the topological organization of tissue structures. We propose ABD-TDA, a deep and topological feature fusion framework that combines transformer-based visual features with complementary topological descriptors. A pretrained ViT extracts a 768-dimensional visual representation, while a dedicated topological branch processes 1,200 Betti-0 and Betti-1 features derived from RGB and HSV color spaces. The resulting 64-dimensional topological representation is fused with the ViT features for classification. We evaluate ABD-TDA on the ICIAR2018 breast cancer and UT-Osteosarcoma bone cancer datasets, containing four and three classes, respectively. The framework achieves 94.10% and 96.54% accuracy, with macro-AUC scores of 99.42% and 99.65%, respectively. Ablation results show consistent improvements from feature fusion, increasing accuracy from 92.67% with ViT alone to 94.10% on ICIAR2018 and from 94.15% to 96.54% on UT-Osteosarcoma. These results demonstrate that topological descriptors provide complementary structural information to transformer-based visual features, leading to more discriminative representations for histopathological cancer classification.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Vision Transformer, Histopathology, Classification, Cancer, Persistence Homology.

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_bx3rOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ABD-TDA, a framework that combines ViT features with topological features for histopathological cancer classification. The ViT branch uses the pretrained google/vit-base-patch16-224 model to extract a 768-dimensional representation. In parallel, the TDA branch computes Betti-0 and Betti-1 curves from the six RGB and HSV channels, resulting in a 1,200-dimensional topological descriptor, which is subsequently reduced to 64 dimensions using an MLP. The two representations are concatenated and passed to an MLP classifier. The proposed fusion improves classification performance over the ViT-only baseline on both ICIAR2018 and UT-Osteosarcoma datasets.

Critical strengths
  1. The paper is easy to follow, and the components used in the paper is simple and easy to understand.
  2. The ablation study is useful because it separately evaluates the TDA-only, ViT-only, and fused models. The fused representation consistently improves over the ViT-only baseline on both datasets and across multiple metrics.
Critical weaknesses
  1. The motivation for introducing TDA is not clear. The paper argues that TDA provide complementary information about tissue organization but it does not clearly explain which biologically or diagnostically meaningful structures in these two cancer-classification tasks are expected to be captured by these quantities. Additionally, the arguement that histopathology contains “structural information” is too generic.

  2. The paper uses the ImageNet-pretrained google/vit-base-patch16-224 as its visual encoder. However, the paper itself discusses recent pathology-specific foundation models such as UNI, CONCH, and Virchow, which are pretrained on large-scale histopathology data and are specifically designed to provide transferable pathology representations. Given the availability of these considerably more relevant backbones, the choice of a generic ViT-Base requires justification.

  3. The paper title suggests the model as attention-based, while the only attention mechanism in ABD-TDA appears to be the standard multi-head self-attention already present in the pretrained ViT backbone.

  4. Table 2 is confusing because different methods use different train test splits. If the reported numbers are taken directly from prior papers, then several methods are evaluated using different train/test splits and potentially different test sets. In that case, the reported accuracies are not directly comparable

  5. The methodological novelty appears limited. The proposed framework combines existing components. While the empirical combination may be useful, it is unclear whether the technical contribution is sufficiently substantial for ICLR

Questions for authors
  1. What histopathological structures do the Betti-0 and Betti-1 features correspond to in these datasets? Can the authors provide qualitative examples showing how differences in the Betti curves relate to diagnostically meaningful tissue morphology?

  2. Why were Betti features extracted independently from RGB and HSV intensity channels? Is there evidence that the topology of these individual color channels reflects meaningful tissue organization rather than primarily stain/intensity variation?

  3. Why was google/vit-base-patch16-224 selected as the visual backbone?

Reviewer_qBY4OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ABD-TDA, combining Vision Transformer (ViT) features with persistent-homology descriptors extracted from RGB and HSV channels for histopathological cancer classification. Experiments on BACH/ICIAR2018 and UT-Osteosarcoma report 94.10% and 96.54% accuracy, respectively. Ablations show improvements of 1.43 and 2.39 percentage points over ViT alone.

Critical strengths

Simple and interpretable architectures. The combination of persistent-homology descriptors and transformer representations is straightforward, and the evaluation across two cancer datasets provides preliminary evidence of applicability.

Critical weaknesses
  1. Unclear evaluation validity and reproducibility. The paper states an 80:20 BACH split, but Appendix Table 4 reports 13,189 training and 5,659 test samples, corresponding to approximately 70:30. Moreover, the original 400 BACH images become 18,848 samples without sufficient explanation of patch generation or source-image-level splitting. Similar provenance concerns exist for UT-Osteosarcoma. Potential leakage cannot be ruled out, and the validation procedure for early stopping is unspecified.
  2. Limited novelty and insufficient experimental validation. Related work already combines persistent-homology features with deep representations, including a closely related TDA–ViT fusion architecture. The paper lacks targeted RGB/HSV and Betti-0/Betti-1 ablations to establish the contribution of its specific topological representation. Results also lack repeated-run uncertainty estimates, while comparisons with prior methods use potentially different evaluation protocols.
Questions for authors
  1. How were the reported BACH and UT-Osteosarcoma samples generated, and were patches from the same original image, slide, or patient separated across training and testing? Why does the BACH split ratio disagree with Appendix Table 4?
  2. Which dataset and metric were used for early stopping and checkpoint selection? Was the test set kept entirely independent of model selection?
  3. What is the specific methodological novelty relative to existing TDA–ViT fusion methods? Could the authors provide RGB-only, HSV-only, and Betti-0/Betti-1 ablations to isolate the proposed contribution?
  4. Are the reported improvements consistent across multiple random seeds or independent splits? Could the authors provide mean performance and uncertainty estimates?
Reviewer_UofBOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ABD-TDA, a feature-fusion framework for histopathological cancer classification. The method combines a pretrained ViT-Base/16 model with handcrafted topological descriptors extracted from RGB and HSV image channels. Persistent homology is used to compute Betti-0 and Betti-1 curves, resulting in a 1,200-dimensional feature vector. A multilayer perceptron reduces these descriptors to 64 dimensions, which are concatenated with the 768-dimensional ViT representation and passed to a classification head.

The framework is evaluated on the BACH breast cancer and UT-Osteosarcoma datasets, achieving classification accuracies of 94.10% and 96.54%, respectively. Ablation experiments report improvements over the ViT-only and TDA-only configurations.

Critical strengths

The paper investigates a reasonable combination of learned visual representations and explicit topological descriptors. The motivation is understandable, and the ablation results provide preliminary evidence that Betti-based features can contribute complementary information to a pretrained ViT on the two selected datasets. The framework is straightforward to implement and computationally modular. The authors describe the feature extraction and fusion procedures clearly and provide an anonymous code repository, which could facilitate reproduction of the proposed approach.

Critical weaknesses
  1. The methodological novelty is insufficient for ICLR. The proposed method mainly concatenates pretrained ViT features with established Betti-curve descriptors processed by an MLP. It does not introduce a new attention mechanism, topological representation learning technique, or feature interaction strategy. Furthermore, related work has already explored topological descriptors for histopathological classification, and the paper does not sufficiently establish a substantive methodological advance over these approaches.

  2. The experimental validation has substantial limitations. The comparisons primarily involve older CNN and handcrafted-feature methods, without modern pathology foundation models such as UNI, CONCH, or Virchow. More importantly, the BACH experiment uses patches derived from a substantially smaller original image collection, but the paper does not clearly establish whether the train/test split prevents overlap between patches originating from the same image or patient. The results also lack repeated-run statistics and significance testing. These issues limit confidence in the reported performance improvements.

Questions for authors
  1. What is the specific methodological contribution of ABD-TDA beyond concatenating existing persistent-homology descriptors with pretrained ViT features? In particular, how does the proposed framework differ substantively from earlier deep–topological fusion approaches?

  2. How were the patch-level datasets partitioned? For BACH, were all patches originating from the same original microscopy image assigned exclusively to either training or testing? Please clarify the original-image or patient-level separation and provide the corresponding split protocol.

  3. Have the authors compared ABD-TDA against modern pathology foundation models, such as UNI or CONCH, under matched training and testing conditions? Additionally, does the TDA branch provide consistent gains when integrated with these stronger pathology-specific visual representations?

  4. Are results from multiple independent training runs available for the ViT-only and ViT+TDA models? Reporting mean performance, standard deviations, and paired statistical comparisons would help establish whether the reported improvements are robust.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.