Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Frequency-based loss reweighting, the standard remedy for imbalanced histopathology segmentation, implicitly assumes that rare classes are difficult. Yet difficulty also arises from morphological variability, boundary ambiguity, and contextual similarity, all largely orthogonal to class frequency. We propose Dynamic Focal Attention (DFA), a simple, efficient mechanism that learns class-specific difficulty directly within the cross-attention of query-based mask decoders. DFA adds a learnable per-class bias to the attention logits, reweighting representations before prediction rather than gradients after it. Initialised from a centred log-frequency prior to prevent gradient starvation and then optimised end-to-end, the bias adapts to difficulty signals as they emerge during training, unifying frequency- and difficulty-aware reweighting in a single attention-bias framework. On three benchmarks (BCSS, BDSA, CRAG), DFA consistently improves Dice and IoU, matching or exceeding a two-stage difficulty-aware baseline without a separate estimator or extra training stage. This shows that encoding difficulty at the representation level is a principled alternative to conventional loss reweighting.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6050_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lakmali240/DFA-Imbalance

Link to the Dataset(s)

BCSS dataset: https://drive.google.com/drive/folders/1BUPZz3nB52J5zRs1ZcEvNK03zw18BeLN?usp=sharing CRAG dataset: https://drive.google.com/drive/folders/1BUPZz3nB52J5zRs1ZcEvNK03zw18BeLN?usp=sharing

BibTex

@InProceedings{RatLak_Learning_MICCAI2026,
        author = { Rathukohe Mudiyanselage, Lakmali Nadeesha Kumari AND Cheung, Sen-Ching Samson},
        title = { { Learning Class Difficulty via Dynamic Focal Attention for Histopathology Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes Dynamic Focal Attention, a learnable attention-based mechanism that models class difficulty dynamically and reweights feature aggregation accordingly, going beyond frequency-based or fixed focal loss approaches. They test their approach using Path-SAM on 3 pathology datasets

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The redefinition of class difficulty beyond frequency is insightful and well-supported

    -The proposed Dynamic Focal Attention is a simple yet novel formulation that introduces learnable, difficulty-aware bias directly in attention (representation-level).

    • Training time reduction is very applealing
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    • The method is only validated on SAM-Path / query-based transformer segmentation, so its effectiveness beyond this family of architectures remains unproven.

    • The method is only validated on histopathology datasets, so its generalization to other medical imaging modalities is not demonstrated.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Despite some limitations, the paper presents a novel and well-motivated approach to modeling class difficulty beyond frequency.

    The method is simple, effective, and consistently improves performance, making it an interesting and worthwhile contribution.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper addresses the problem of class imbalance in semantic segmentation of histopathology images. Typical methods use approaches that focus on class frequency to modify weights accordingly, assuming that rare classes are the most difficult ones. The authors show that this is not always true, and morphological variability, boundary ambiguity and other factors have a strong impact in class difficulty. They propose a method that, building on SAM-Path semantic segmentation model, introduces class difficulty as a bias scalar term per class in the focal attention calculation. Three approaches for the bias calculation are compared: using pixel frequency of each class, using a Dice coefficient per class obtained with a pre-trained baseline model, and a dynamic alternative that treats the bias as a learnable scalar that is optimized during training, which is the main proposal of the authors. An initialization strategy for the learnable scalar based on class frequency is also proposed as a more robust alternative to the zero-initialisation. It is shown that a higher scalar value corresponds with a badly segmented class, while a lower value corresponds with better segmentation. The method is evaluated on three public histopathology image databases. The different attention bias proposals are compared on the three benchmarks. Results show that the proposed dynamic focal attention method outperforms the rest for most of the classes in terms of Dice and IoU, especially improving the most difficult classes.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    It presents a generic histopathology image semantic segmentation model, which is a clinically relevant problem, that shows robust performance on three different databases with multiple classes. It addresses the class imbalance problem in segmentation of histopathology images with an original approach that adds little computational cost to the baseline method. It proposes a dynamic focal attention method to improve the segmentation of especially difficult classes, and the authors empirically show that there is a significance improvement in those cases on different classes and databases.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Performance improvement with DFA is consistent but small. In general, there is a bigger improvement from Base w/o FL to CFFA, but very little from CFFA to the proposed DFA. It is true that it does not add a lot of computational complexity with respect to CFFA, but it does not bring a lot of performance improvement either. A similar thing occurs with the proposed warm initialization of the learnable bias (Table 2): there is a big difference in the learned scalars between cold and warm initialization, but this does not reflect so much in a Dice improvement. There may be classes that are difficult in some of the images but easier in other images, such as tumour or stroma, what would happen with the learnt difficulty score in those cases? Would it affect the attention and thus the final segmentation of the class? Intra-class inhomogeneity is not addressed in any way. Training strategy:

    • The compared methods are trained for few epochs in each database (30, 20 and 60). Why those values? With more epochs, would DFA, CFFA and HCFA performance results be more similar? Would there still be a consistent improvement with the proposed DFA strategy? From my point of view, this is a point of concern.
    • Experimental setup: how many images are there in each database? A few more details would be helpful. Also, it is not mentioned how train/validation/test are organized in the training strategy. Is one split used? A cross validation scheme? Minor things:
    • Section 2.2 should belong to the results section since it is not a methodological description but rather an analysis between class pixel frequency and Dice score on the databases
    • At the end of the method description, an Lreg term is mentioned but it is not clear where it belongs since it does not appear anywhere in equation 7.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    It presents a novel way of learning class difficulty during training and shows it consistently improves semantic segmentation in three different histopathology image databases with little additional computational cost. However, the improvement is rather small and there are some concerns about the training strategy that would need clarification. Besides, the contribution seems to be incremental over previous works.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors address all my concerns in a satisfactory way. I believe the revised manuscript should be accepted for MICCAI.



Review #3

  • Please describe the contribution of the paper

    This work addresses semantic segmentation in histopathology. Even though the title includes “Imbalanced Histopathology Segmentation”, the authors tackle the difficulty of histopathology segmentation, showing that the frequencies and segmentation difficulties of each class are disconnected. They think difficulties stem from morphological variability, boundary ambiguity and contextual similarity. The authors proposed three variants of the attention mechanism for the SAM-Path decoder: Class Frequency Focal Attention (CFFA), Hard Class Focal Attention (HCFA), and Dynamic Focal Attention (DFA), using a Focal Attention Bias framework. They present experimental results using three publicly available segmentation datasets: BCSS (breast cancer), BDSA (brain tissue) and GRAG (colorectal adenocarcinoma).

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    -Explore the difficulty of medical segmentation -Proposal of three attention mechanisms under a Focal Attention Bias framework -Experimental evaluation with three publicly available datasets of different parts: breast, brain and colon.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    -Misleading title -Insufficient survey of related works -Incorrect mathematical definitions and unclear derivation -Unclear explanations -Incorrect visual explanation of Fig. 1.-Unconvincing experimental evaluations without comparative evaluations

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The title is misleading. If the authors focus on the difficulty of medical segmentation, they should give a more appropriate title. Furthermore, if the authors focus on the difficulties stemming from morphological variability, boundary ambiguity and contextual similarity, they should present related work from their viewpoint. There are many previous works, such as edge-aware, boundary-aware, boundary-ambiguity-aware, and context-aware methods. Eq. (1) is mathematically incorrect. Since \bm{QK}^{\top} and \bm{b} are a matrix and a vector, the summation of them is undefined in linear algebra. In Eqs. (2), (3), (5) and (6), \cdot is undefined. If the authors want to use it for a scalar multiplication, it should be omitted. The derivation of Eq. (5) is unclear. From Eq. (1), how did the authors introduce Eq. (5)? How to convert the target of partial derivatives? There is no explanation and its validity is unclear. In Eq. (7), \mathcal{L}{dice}, \mathcal{L}{focal}, \mathcal{L}{mse}, IOU and \hat{iou}{c}. The work frequency should be mathematically defined to ensure reproducibility. In Fig. 1, the output Queries, Keys, Values, and Attention logits have inconsistent sizes to illustrate their operations. Furthermore, for the size notation, Fig. 1 uses the alphabet x instead of the operator \times. The operations \otimes and \oplus in Fig. 1 are also unclear or undefined. The results of Fig. 3 and Fig. 2 are unconvincing for the authors’ claim. The results of Fig. 3 show a correlation to class frequency, while Fig. 2 shows frequency-difficulty disconnection. Furthermore, since the definition and explanation of the proposed method (DFA) are unclear, its actual efficacy is also unclear. In the experiments, no figure shows the obtained attention, which should be compared with related work. Without demonstrating the obtained attention and comparative evaluations, it is unclear whether the proposed method mitigates the difficulties stemming from morphological variability, boundary ambiguity, and contextual similarity.

    As a result, I conclude that the submission is premature for MICCAI presentation.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    The authors’ rebuttal is unconvincing, and the fundamental problems remain.

    For example, even though they claim to refine the title to “Learning Class Difficulty via Dynamic Focal Attention for Histopathology Segmentation,” they do not clarify the difficulty of histopathology images for the segmentation task. Instead, they just say “the primary evidence for the difficulty-learning is the learned class-level attention bias \delta_c profile and per-class performance”. It does not make sense. It seems they focus solely on the performance attack rather than exploring the true class difficulty in histopathology segmentation. Where did the segmentation difficulty come from ? I think the image’s characteristics. Since there is no guarantee that Q, K, and V remain unchanged before and after adding a learnable \delta_c during training, comparative evaluations and visualisations should be necessary for a fair presentation that will clarify the segmentation difficulty in an input image for the histopathology segmentation. I think the experimental evaluations and presentations don’t match their titles.

    Furthermore, the authors’ claims about Figs. 2 and 3 are irrational. In BCSS, Inflammatory is the hardest class in Fig. 3, and the Dice score is the smallest among the five classes in Fig. 2.However, it might come from the low frequency. On the other hand, “Necrosis” is the second hardest class in Fig. 3, but the Dice score is almost the same as the second highest one (”Stroma”) in Fig. 2.In addition, even though “Other” is 0.158 smaller than “Necrosis” in Fig. 3, that should mean “Other” is easier than “Necrosis”; the Dice score of “Other” is 5.6% smaller than “Necrosis”. From the results, I do not think that \delata_c accurately reflects the true difficulty of pathological segmentation. The authors’ analysis lacks fairness and objectivity.

    Moreover, given the current status, I cannot expect satisfactory revisions to definitions, figures, descriptions of related work, the logical flow and readability. Even if they can, the camera-ready manuscript should become a fully different manuscript from the submitted one. NOTE: If the authors think reproducibility is important for the scientific work, the correct derivation of Eq. 5 with appropriate description and explanations should have been presented, even if it was in the Appendix or supplemental material.



Author Feedback

We thank reviewers for constructive feedback. We are encouraged that R1/R2 found the method novel, simple, and consistently effective, and we will revise the paper to improve clarity, reproducibility, and positioning.

R1 — Generalization. Focal Attention Bias requires only cross-attention with class queries, so it is directly applicable to other query-based mask decoders. Cross-architecture and cross-modality validation is planned for a journal extension. R2 — CFFA→DFA gap. The mean gain is modest because CFFA already captures the frequency-related component of difficulty; DFA targets the non-frequency residual, so gains concentrate on hard classes: Inflammatory +2.7, Other +0.7, Leptomeninges +0.7, Superficial +0.5, Gland +1.7 Dice, with no easy-class tradeoff. We will emphasize these per-class difficulty-sensitive gains. R2 — Warm-start. Cold initialization can keep $\delta_c$ near zero because rare classes receive weak gradients through the softmax self-gating term. Warm-start avoids this collapse by yielding the intended class-discriminative profile by boosting/lowering $\delta_c$ for hard/easy classes. Thus, warm-start is useful not only for the +0.75 Dice gain, but also for stable learning of the difficulty profile. R2 — Intra-class inhomogeneity. DFA models class-level expected difficulty, not image-specific difficulty. Local adaptation is still governed by the token-dependent $QK^T$ term, which preserves spatially varying feature matching. We will clarify this scope and discuss region-conditioned difficulty bias as future work. R2 — Training schedule/splits. The 30/20/60 epoch schedules were selected from validation-loss plateaus and were identical for all compared methods on each dataset; thus, the schedule does not preferentially benefit DFA over CFFA/HCFA. We used one fixed train/val/test split per dataset following the SAM-Path protocol, with 20% of training held out for validation. For BDSA, 10 WSIs produce ~82K 1024^2 patches, with held-out data split at the slide level. We will add these details to Sec. 3.1.R2 — Sec. 2.2 and L_reg. Agreed. Sec. 2.2 is empirical analysis and will move to the Results section; Eq. 7 will explicitly include L_reg. R3 — Title and related work. We will retitle to “Learning Class Difficulty via Dynamic Focal Attention for Histopathology Segmentation.” We will expand related work positioning DFA as a complementary class-level attention-bias intervention. R3 — Equations/notation. Eq. 1 uses broadcast addition of a class-wise bias $b$ across $N$ spatial positions; we will denote this by $\oplus$, define all tensor shapes, and remove unnecessary $\cdot$ notation. Eq. 5 was compressed for space; the full softmax Jacobian is $\partial \tilde{\alpha}{i,k} / \partial \delta_c = \tilde{\alpha}{i,k}(\mathbf{1}{k=c} - \tilde{\alpha}{i,c})$, with on-diagonal self-gating $\tilde{\alpha}{i,c}(1 - \tilde{\alpha}{i,c})$ motivating the gradient-starvation argument. We will include the complete derivation. R3 — Eq. 7 / Fig. 1.We will define L_dice, L_focal, L_mse, f_c, IoU, and iou_c explicitly. Fig. 1 will be redrawn with shape-consistent Q (C x d), K, V (N x d), and logit (C x N) blocks; $\otimes$ defined as matrix multiplication, $\oplus$ as broadcast addition, and x used for size notation. R3 — Fig. 2 vs Fig. 3.The figures are complementary: Fig. 2 shows frequency and difficulty are not reliably aligned, while Fig. 3 relates learned $\delta_c$ to baseline Dice. For example, in BCSS, Inflammatory and Necrosis have near-identical frequency but learned delta = +0.41 vs. +0.07, matching a 22-point Dice gap. This cannot be produced by a purely frequency-driven scheme. R3 — Attention visualization. DFA modulates class-level attention bias, while spatial selectivity remains governed by $QK^T$; thus, the primary evidence for the difficulty-learning claim is the learned $\delta_c$ profile and per-class performance gains, rather than attention-map visualization. Code will be released upon acceptance



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The work proposes a class difficulty learning method for histopathology image segmentation with imbalanced issue by developing a dynamic focal attention mechanism. The recommendation is mixed among reviewers. The main disagreements are on empirical evaluation, methodology effectiveness, lack of sufficient evidences and explanations for some claims, and some writing issues. Based on my reading, I think that the reviewers raised effective concerns. Please try to fix these concerns in the rebuttal.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Two reviewers agree to accept and one reviewer for rejecting. I read the rebuttal and reviewers’ comments after the rebuttal. I think that the rebuttal solves the main concerns, though it does not fully eliminate the concerns. But due to the novelty of the method, I tend to accept.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    After rebuttal, this paper receives two accept recommendations and one reject recommendation. The AC acknowledges that the paper addresses an important issue in imbalanced histopathology segmentation and proposes a simple dynamic focal attention mechanism to learn class difficulty beyond frequency, with consistent but modest gains across three public datasets. However, the rebuttal does not sufficiently resolve the major concerns about the clarity and rigor of the method formulation, the correctness and completeness of mathematical definitions, the strength of the evidence supporting the claimed interpretation of learned class difficulty, and the lack of convincing comparative analysis against related boundary-, context-, or difficulty-aware segmentation methods. Several key revisions would substantially change the submitted manuscript rather than merely clarify it. Therefore, despite the interesting motivation and positive reviews, the AC believes that the current paper remains methodologically and experimentally under-supported, and recommends rejection.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The majority of reviewers find this paper insightful. I also consider the idea of integrating class-specific bias into the attention mechanism to be both novel and inspiring. I agree in part with R3’s comments that the paper could be better organized and that some methodological details are missing. The authors are encouraged to include more method details and ablation studies in a future version.



back to top