Pith. sign in

REVIEW 4 major objections 5 minor 20 references

Recurrent Contrastive Learning for Imbalanced Medical Image Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Recurrent anchors expand tail-class regions in medical imaging

desk verdict A novel recurrent anchor-field mechanism for tail-class support expansion, but the empirical validation lacks statistical rigor and the gains are backbone-dependent. read the letter →

arxiv 2608.03304 v1 pith:H2AZVPAJ submitted 2026-08-04 cs.CV

Zhiyuan Zhu , Xinling Meng , Junxuan Yu , Jiongquan Chen , Qiongying Ni , Tuhang Shao , Yuhao Huang , Luping Zhou
show 10 more authors
Ruiyang Huang Yuxue Wang Rongliang Zhang Xue Wang Tianhong Tang Likun Wang Junbo Chen Yong Jiang Yongping Lu Xin Yang
This is my paper
classification cs.CV
keywords imbalancedmedicalimageclassificationcontrastivelearningtemporalmemoryqueueanchorssupportregionexpansionclassimbalanceDINOv3LoRA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes Recurrent Contrastive Learning (RCL), a training framework that treats class-imbalanced medical image classification as a problem of enlarging the latent support region of rare classes. The authors argue that existing reweighting and resampling methods only adjust how observed features are used, leaving tail-class representations compact and vulnerable to head-class encroachment. RCL instead builds an anchor field around tail classes by reusing historical feature states stored in a Temporal Memory Queue, pulling rare-class features toward a larger, more separable region. On three imbalanced medical datasets, RCL improves balanced accuracy over strong baselines, supporting the claim that recurrently constructed temporal anchors counteract the geometric bias of imbalanced data.

What carries the argument

The Temporal Memory Queue (TMQ) is the core object: it stores EMA-updated feature banks from previous epochs, providing a corpus-level, temporally diverse set of references. From TMQ, the method constructs Temporal Anchors (TARs) for each tail class by applying K-means, selecting the bottom-k features farthest from the cluster center, and adding Gaussian noise. These anchors are injected into the recurrent contrastive loss, which resembles a supervised contrastive loss over both current-bank features and anchors. The anchor field is what enlarges the tail-class support region and creates a buffer between head and tail classes.

What would settle it

Replace the bottom-k-and-noise anchor construction with randomly chosen tail-class features or non-noisy centroids. If balanced accuracy does not drop, or if the decision boundary does not shift toward head classes, then the specific anchor-selection rule is not what drives the improvement.

Watch

Extended reading notes

Core claim

The central claim is that enlarging the tail-class support region in feature space, rather than merely rebalancing the loss or sampling, yields better classification under class imbalance. RCL maintains a Temporal Memory Queue (TMQ) of epoch-wise feature banks, each updated with an exponential moving average of projection embeddings. At each training epoch, it selects bottom-k features from tail classes in earlier banks, perturbs them with Gaussian noise to form Temporal Anchors (TARs), and adds these anchors to a contrastive loss as extra positive and negative references. The resulting anchor field occupies the frontier around tail-class clusters, acting as a buffer against head-class featu

Load-bearing premise

The whole field-expansion mechanism rests on assuming that features selected from the bottom-k of a tail class's cluster, plus Gaussian noise, are faithful positive examples whose attraction widens the true support region of that class.

Editorial extensions

If this is right

  • If RCL is correct, imbalanced classification can be improved by explicitly reshaping feature geometry rather than only reweighting losses or resampling data.
  • The recurrent use of historical feature banks suggests that training-stage information carries value beyond the final epoch, enabling progressively refined anchors as training evolves.
  • RCL's improvement over vanilla DINOv3 fine-tuning on all three datasets indicates that the method is additive to strong pretrained backbones.
  • The reported QWK of 0.92 on APTOS 2019 implies that RCL can push medical grading tasks closer to or beyond specialized competition-winning performance.
  • RCL's compatibility with a ResNet50 backbone, where it sometimes beats DINOv3 fine-tuning, suggests the field-expansion mechanism is backbone-agnostic.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One can test whether the bottom-k and Gaussian noise selection is essential by replacing it with random tail-class features or no noise; if performance holds, the exact anchor construction is not the active ingredient.
  • The method's reliance on manual hyperparameters (tail-class count, sample number, noise scale) suggests an adaptive anchor-selection rule could further improve stability and generality.
  • RCL's field-expansion view might extend to other imbalance settings such as few-shot learning or open-set recognition, where support-region boundaries also matter.
  • The t-SNE observation that TARs occupy the intermediate region between two classes hints that anchors may also serve as a regularization mechanism on decision boundaries, not just as contrastive positives.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Recurrent Contrastive Learning (RCL) for imbalanced medical image classification. The method uses a DINOv3 backbone with LoRA adapters, maintains a Temporal Memory Queue (TMQ) of epoch-wise feature banks, and constructs Temporal Anchors (TARs) from the bottom-k tail-class features of historical banks, perturbed by Gaussian noise. These anchors are inserted into a contrastive loss so that tail-class features are attracted toward an expanded support region. Experiments are reported on one private carotid ultrasound dataset and two public datasets (APTOS 2019, KneeOA), with balanced accuracy, F1, accuracy, and QWK metrics. The authors claim the best balanced accuracy on all three datasets, outperforming several recent imbalanced-learning baselines.

Significance. If the central claim is substantiated, the anchor-field perspective is a plausible and potentially useful addition to imbalanced medical image classification. The paper contributes a concrete mechanism (TMQ + TARs) and releases code, which supports reproducibility. The use of two public datasets and a private multi-center dataset is a strength. However, the current evidence is not sufficient to establish the claimed generality: there are no error bars or significance tests, hyperparameters appear to be selected on the Carotid dataset after inspecting results, the key assumption in Eq. (4) is untested, and the same-backbone comparison partially contradicts the 'consistent improvements' claim. The method is interesting but needs additional validation before the headline conclusions can be accepted.

major comments (4)
  1. [§3, Table 1] The headline claim of 'best Bacc on all three datasets' rests entirely on point estimates. No standard deviations, confidence intervals, or significance tests are reported, despite the text stating that four-fold cross-validation was performed. On APTOS 2019 the DINOv3→DINOv3+RCL gain is 68.48→69.01 (0.53 Bacc), which is within typical fold-to-fold variability. Without error bars or a paired significance test, the central claim is not established. Please report the per-fold results and a paired comparison against the closest baselines.
  2. [§3, Table 2 and implementation paragraph] The final hyperparameters (sample number n and Tail Classes) are selected based on results on the Carotid dataset, and the text reports only configurations showing 'most significant influence' rather than a pre-specified validation procedure. It is unclear whether the same n and Tail Classes were fixed a priori for APTOS 2019 and KneeOA, or whether the held-out test split was used to choose them. If the reported test numbers informed hyperparameter selection, the comparison is optimistic. Please specify a validation-based selection protocol and report results for fixed hyperparameters across all datasets.
  3. [§2.3, Eq. (4)] The core mechanism assumes that bottom-k features (those farthest from the K-means cluster center) plus Gaussian noise ε∼N(0,0.1²) are valid tail-class positives whose attraction expands the true support region. This assumption is load-bearing and untested. If those features are outliers or off-manifold, the contrastive loss will pull tail-class embeddings toward misleading regions and can degrade the decision boundary. Please provide evidence that TARs are semantically valid: for example, measure nearest-neighbor precision of anchors, compare with anchors centered at cluster means or with random in-class features, and report sensitivity to the noise scale and to k_t.
  4. [§3, Table 1, RCL* row] The claim of 'consistent improvements over strong baselines' is contradicted by the same-backbone comparison. On KneeOA, RCL* (66.65 Bacc) is below BCL* (69.58) even though both use ResNet50. On APTOS 2019, RCL* (66.23) is below the DINOv3 baseline (68.48). In addition, most baselines use ResNet50 while the top RCL row uses DINOv3, so the headline gains are not attributable to RCL alone. Please provide a controlled comparison with the same backbone for all methods, or qualify the claim accordingly.
minor comments (5)
  1. [§2.3, Eq. (4)] The notation 'Bottom-K' is overloaded: it is unclear whether it refers to the number of clusters in K-means or the number of selected features. Please clarify that it selects the k_t features with the lowest similarity to their cluster center.
  2. [§2.2] The term 'recurrent' is used for reusing historical banks, but no recurrent network is involved. Consider clarifying the terminology to avoid confusion with recurrent architectures.
  3. [§2.2] The memory cost of storing M+1 full feature banks is not discussed. With 100 epochs and large datasets this can be nontrivial; please state the actual storage overhead and any trade-offs.
  4. [§3, Table 2] The selection among multiple metrics is ambiguous: the setting with Tail Classes=2 and n=7 gives the best Bacc/F1/Acc, but the highest QWK is obtained with Tail Classes=1 and n=3. State the criterion used to choose the final configuration.
  5. [§3, Quantitative analysis] The comparison with the APTOS 2019 leaderboard (QWK=0.92 vs 0.91) is informal because the challenge used different data splits and training protocols. It should be presented as an external reference, not as a controlled comparison.

Circularity Check

1 steps flagged · score 3.0 of 10

Core empirical claim is independent, but one qualitative 'support expansion' observation is guaranteed by the TAR construction itself.

  1. self definitional [Section 2.3 Eq. (4); Section 3 Qualitative analysis / Fig. 3]
    "These selected features are then perturbed with Gaussian noise as in Eq. 4: T AR(m−st) =Bottom-K(k−means({(q (m−st) i , yi)|y i ∈ T })) + ε∼N(0,0.1 2). ... TARs display a larger radius than current epoch features, suggesting an expanded tail-class boundary induced by the anchor field."

    Bottom-K selects the intra-class features farthest from their K-means cluster center, and adding Gaussian noise with std 0.1 further inflates the spread. Therefore the TAR set is constructed to have a larger radius than ordinary current-epoch features by definition. The paper presents this larger radius as empirical evidence that the anchor field expands the tail-class boundary, but the expansion is baked into Eq. 4 and is not an emergent effect of training. This is a self-definitional validation of the mechanism. It does not affect the held-out Bacc comparisons, which remain independent and ground the main empirical claim.

full rationale

The paper's central claim is empirical: RCL achieves the best balanced accuracy on three imbalanced medical datasets, and these numbers are obtained on held-out test data. The method is not a derivation whose output is equivalent to its input. No parameter is fitted on a subset and then reported as a prediction of the same subset: the hyperparameters in Table 2 are used to choose n and Tail Classes on the Carotid dataset, but the final comparisons are on separate test folds and two other datasets, and the headline results do not reduce to those chosen hyperparameters by construction. The only self-citation is reference [20], used as 'inspired by' for the temporal memory queue; it is not load-bearing because the queue design is explicitly described and is an architectural choice, not a uniqueness theorem or an unverified premise. The one identifiable circular step is the qualitative claim in Fig. 3 that TARs have a larger radius than current-epoch features, which is guaranteed by the bottom-k selection plus Gaussian perturbation in Eq. 4. This is a self-definitional observation, not an independent confirmation of the field-expansion mechanism. However, because the main result is the held-out classification improvement, this circularity is partial and does not undermine the paper's central empirical contribution. The conclusion's admission of manual hyperparameter tuning is a limitation, not a circularity. Overall score 3 reflects the minor self-definitional element while recognizing the independent external grounding of the headline numbers.

Assumptions & free parameters 7 free parameters · 4 assumptions · 2 invented entities

The method's performance depends on a set of hand-chosen hyperparameters and on assumptions about pretrained feature transferability and anchor validity; none are derived or externally benchmarked.

free parameters (7)
  • EMA momentum lambda = 0.01
    Set by hand, controls how much current embeddings update the memory bank; central to temporal memory queue (Eq. 2).
  • contrastive temperature beta = 0.1
    Set by hand, scales logits in contrastive loss (Eq. 6).
  • loss weight alpha = linearly decayed from 1.0 to 0.1
    Balances cross-entropy and contrastive loss; schedule chosen by hand.
  • sample step s = 3
    Interval between sampled historical epochs for anchor construction.
  • sample number n = 7 in best setting
    Number of historical banks used for anchors; chosen via experiments in Table 2.
  • bottom selection ratio k_t = 50% to 10% decreasing
    Proportion of farthest-from-center features selected as anchors; schedule chosen by hand.
  • number of tail classes T = 2 in best setting
    Which tail classes receive anchors; selected based on validation results.
assumptions (4)
  • domain assumption DINOv3 frozen embeddings are strong enough that only LoRA adapters need training
    Section 2.1 freezes the backbone; if the pretrained features are not transferable, the whole pipeline loses its foundation.
  • domain assumption Historical feature banks from earlier epochs remain informative and aligned with the current representation
    EMA updates (Eq. 2) assume stored features track the evolving space; no analysis of drift is provided.
  • domain assumption K-means cluster centers are a valid summary of tail-class feature structure
    Used to select bottom-k anchors in Eq. 4; assumes the clusters are meaningful.
  • ad hoc to paper Gaussian noise with std 0.1 creates valid on-manifold neighbors
    Eq. 4 adds epsilon ~ N(0, 0.1^2); no justification that the perturbed embeddings remain semantically valid.
invented entities (2)
  • Temporal Memory Queue (TMQ)
    purpose: Preserve corpus-level features from multiple training epochs for anchor construction
    Algorithmic construct introduced in Section 2.2; its benefit is only shown through the paper's own experiments.
  • Temporal Anchors (TARs)
    purpose: Expand tail-class support region in latent space during contrastive learning
    Constructed in Eq. 4 from historical bank features; no external validation of the field-expansion mechanism.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recurrent Contrastive Learning for Imbalanced Medical Image Classification." pith.science (2026). https://pith.science/paper/H2AZVPAJ

@misc{pith2026260803304,
  author       = {Pith},
  title        = {Pith review of: Recurrent Contrastive Learning for Imbalanced Medical Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H2AZVPAJ}},
  note         = {Machine review of arXiv:2608.03304}
}
read the original abstract

Medical image classification often suffers from class imbalance due to the inherent disparities in disease incidence. Existing approaches, such as class resampling and loss reweighting, mainly improve learning within the observed feature distribution, but do not explicitly enlarge the latent support region of tail classes. As a result, tail-class representations remain overly compact and are easily encroached upon by head classes, leading to biased decision boundaries. In this work, we propose Recurrent Contrastive Learning (RCL) for imbalanced medical image classification. RCL progressively expands the support region of tail classes by recurrently reusing historical feature states across training phases. Specifically, we adopt DINOv3 with LoRA adapters as the backbone to provide robust feature embeddings. We then devise a Temporal Memory Queue (TMQ) to preserve corpus-level features across training phases and provide diversified global references for contrastive learning. Based on TMQ, we construct Temporal Anchors (TARs) to form an anchor field around tail classes. This field enlarges the support region of tail classes, suppresses head-class encroachment, and improves inter-class separation. Extensive experiments on three imbalanced medical datasets demonstrate that RCL achieves consistent improvements over strong baselines. The code is available at https://github.com/dndins/RCL.

Figures

Figures reproduced from arXiv: 2608.03304 by the authors.

Figure 1
Figure 1. Comparison of different imbalanced classification strategies and data distribu￾tions. (a): Data imbalance situation; (b-c): Data reweighting and resampling; (d): our anchor-based RCL; (e-g): The class distribution of three datasets To address this class-imbalanced challenge, various strategies have been pro￾posed in recent years. Menon et al. [8] introduced a logit adjustment strategy that incorporates class-prior i… view at source ↗
Figure 2
Figure 2. Overview of the proposed RCL framework. 2.1 DINOv3 with LoRA Adapter Fine-tuning The foundation model acquires strong representation capability through self￾supervised learning on large-scale data, which enables effective adaptation to downstream domains with limited labeled data. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. t-SNE visualization and Bacc evolution at different training stages. eter configurations that show the most significant influence on the experimental results. Here, Sample Number n denotes the number of historical memory banks sampled to construct anchors. Tail Classes denotes the number of tail classes involved as TARs. The setting with Tail Classes = 2 and n = 7 achieves the best overall performance, yielding the … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

  1. [1]

    Knee osteoarthritis severity grading dataset.Mendeley Data, 1(10.17632):30784984, 2018

    Pingjun Chen. Knee osteoarthritis severity grading dataset.Mendeley Data, 1(10.17632):30784984, 2018

    work page 2018

  2. [2]

    Dy- namic feature splicing for few-shot rare disease diagnosis.Medical Image Analysis, 90:102959, 2023

    Yuanyuan Chen, Xiaoqing Guo, Yongsheng Pan, Yong Xia, and Yixuan Yuan. Dy- namic feature splicing for few-shot rare disease diagnosis.Medical Image Analysis, 90:102959, 2023

    work page 2023

  3. [3]

    Deepsmote: Fusing deep learning and smote for imbalanced data.IEEE Transactions on Neural Net- works and Learning Systems, 34(9):6390–6404, 2022

    Damien Dablain, Bartosz Krawczyk, and Nitesh V Chawla. Deepsmote: Fusing deep learning and smote for imbalanced data.IEEE Transactions on Neural Net- works and Learning Systems, 34(9):6390–6404, 2022

    work page 2022

  4. [4]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR, pages 770–778, 2016

    2016

  5. [5]

    Hifuse: Hierarchical multi-scale feature fusion net- work for medical image classification.Biomedical Signal Processing and Control, 87:105534, 2024

    Xiangzuo Huo, Gang Sun, Shengwei Tian, Yan Wang, Long Yu, Jun Long, Wen- dong Zhang, and Aolun Li. Hifuse: Hierarchical multi-scale feature fusion net- work for medical image classification.Biomedical Signal Processing and Control, 87:105534, 2024

    work page 2024

  6. [6]

    Aptos 2019 blindness detection dataset.https://www

    M Karthick and D Sohier. Aptos 2019 blindness detection dataset.https://www. kaggle.com/competitions/aptos2019-blindness-detection, 2019

    work page 2019

  7. [7]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InCVPR, pages 2980–2988, 2017. 10 Z. Zhu et al

    work page 2017

  8. [8]

    Long-tail learning via logit adjustment.arXiv preprint arXiv:2007.07314, 2020

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment.arXiv preprint arXiv:2007.07314, 2020

    arXiv 2007

Show all 20 references
  1. [9]

    Equilibrium contrastive learning for imbalanced image classification.arXiv preprint arXiv:2602.09506, 2026

    Sumin Roh, Harim Kim, Ho Yun Lee, and Il Yong Chun. Equilibrium contrastive learning for imbalanced image classification.arXiv preprint arXiv:2602.09506, 2026

    2026

  2. [10]

    Carotid plaque-rads: a novel stroke risk classification system

    Luca Saba, Riccardo Cau, Alessandro Murgia, Andrew N Nicolaides, Max Win- termark, Mauricio Castillo, Daniel Staub, Stavros K Kakkos, Qi Yang, Kosmas I Paraskevas, et al. Carotid plaque-rads: a novel stroke risk classification system. Cardiovascular Imaging, 17(1):62–75, 2024

    2024

  3. [11]

    Handling imbalanced medical datasets: review of a decade of research.Artificial Intelligence Review, 57(10):273, 2024

    Mabrouka Salmi, Dalia Atif, Diego Oliva, Ajith Abraham, and Sebastian Ventura. Handling imbalanced medical datasets: review of a decade of research.Artificial Intelligence Review, 57(10):273, 2024

    2024

  4. [12]

    Similarity memory prior is all you need for medical image segmentation

    Hao Tang, Zhiqing Guo, Liejun Wang, and Chao Liu. Similarity memory prior is all you need for medical image segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 23009–23018, 2025

    2025

  5. [13]

    Disease-grading networks with asymmetric gaussian distribution for medical imaging.IEEE Transactions on Medical Imaging, 2025

    Wenqiang Tang and Zhouwang Yang. Disease-grading networks with asymmetric gaussian distribution for medical imaging.IEEE Transactions on Medical Imaging, 2025

    2025

  6. [14]

    Mixture-of-experts learner for single long-tailed domain generalization

    Mengzhu Wang, Jianlong Yuan, and Zhibin Wang. Mixture-of-experts learner for single long-tailed domain generalization. InACM MM, pages 290–299, 2023

    2023

  7. [15]

    Proco: Prototype-aware contrastive learning for long-tailed medical image classification

    Zhixiong Yang, Junwen Pan, Yanzhan Yang, Xiaozhou Shi, Hong-Yu Zhou, Zhicheng Zhang, and Cheng Bian. Proco: Prototype-aware contrastive learning for long-tailed medical image classification. InMICCAI, pages 173–182. Springer, 2022

    2022

  8. [16]

    Adaptive dual-axis style-based recalibration network with class-wise statistics loss for imbalanced medical image classification

    Xiaoqing Zhang, Zunjie Xiao, Jingzhe Ma, Xiao Wu, Jilu Zhao, Shuai Zhang, Runzhi Li, Yi Pan, and Jiang Liu. Adaptive dual-axis style-based recalibration network with class-wise statistics loss for imbalanced medical image classification. IEEE Transactions on Image Processing, 2025

    2025

  9. [17]

    Ecl: Class-enhancement contrastive learning for long-tailed skin lesion classification

    Yilan Zhang, Jianqi Chen, Ke Wang, and Fengying Xie. Ecl: Class-enhancement contrastive learning for long-tailed skin lesion classification. InMICCAI, pages 244–254. Springer, 2023

    2023

  10. [18]

    Learning from neighbors: Category extrapolation for long-tail learning

    Shizhen Zhao, Xin Wen, Jiahui Liu, Chuofan Ma, Chunfeng Yuan, and Xiaojuan Qi. Learning from neighbors: Category extrapolation for long-tail learning. In CVPR, pages 30483–30492, 2025

    2025

  11. [19]

    Balanced contrastive learning for long-tailed visual recognition

    Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. InCVPR, pages 6908–6917, 2022

    2022

  12. [20]

    Hierarchical corpus-view- category refinement for carotid plaque risk grading in ultrasound

    Zhiyuan Zhu, Jian Wang, Yong Jiang, Tong Han, Yuhao Huang, Ang Zhang, Kai- wen Yang, Mingyuan Luo, Zhe Liu, Yaofei Duan, et al. Hierarchical corpus-view- category refinement for carotid plaque risk grading in ultrasound. InMICCAI, pages 255–264. Springer, 2025

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.