List of Papers Browse by Subject Areas Author List
Abstract
High-resolution medical images contain fine-grained, spatially sparse cues that are critical for diagnosis, yet preserving full resolution incurs substantial computational and memory costs. Most deep models process images uniformly, leading to redundant computation or loss of diagnostic detail under downsampling. We propose Chained Perceptual Refinement (CPR), a coarse-to-fine framework that formulates medical image analysis as a sequential global-to-local decision process. Starting from a low-resolution global view, CPR dynamically predicts the center and spatial extent of refinement regions, extracts high-resolution evidence from the original image, and incrementally integrates it with global context. By keeping the backbone input size fixed while contracting the perceptual field, CPR preserves diagnostic fidelity with constant peak GPU memory. Extensive experiments on five medical imaging datasets and multiple backbone architectures demonstrate that CPR consistently outperforms both fixed-resolution and multi-scale state-of-the-art baselines, achieving improvements of up to 2.27% over the second-best method. It also achieves up to a 19.6× reduction in GFLOPs at matched accuracy, establishing a superior accuracy–efficiency trade-off for high-resolution medical image analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3366_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LuSi_CPR_MICCAI2026,
author = { Lu, Si-Yuan AND Zhu, Hanruo AND Zhu, Ziquan AND Jin, Gaojie AND Fu, Zeyu AND Yin, Lu AND Li, Ke AND Liu, Lu AND Huang, Tianjin},
title = { { CPR: Chained Perceptual Refinement for Coarse-to-Fine Medical Image Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a CPR framework that reconstructs high-resolution medical image classification as a sequential decision-making process from global to local. By dynamically predicting the center coordinates and spatial range of the refined region at each step, adaptive perceptual field contraction is achieved while keeping the input size of the backbone network fixed, thus achieving a balance between diagnostic accuracy and computational efficiency; This method has achieved advanced performance on multiple publicly available medical imaging datasets of different modalities
- 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.
This paper (1) innovatively reconstructs the classification of high-resolution medical images as a sequential decision-making process from global to local. The strategy network simultaneously predicts the center coordinates and spatial range of the refined region, achieving dynamic scale contraction, which is closer to the gradual focus diagnosis process of clinical doctors than the fixed size and unconstrained block sampling in existing methods. (2)The state transition mechanism accumulates evidence in a residual manner without increasing input resolution. (3) At the same time, a reduction in computational complexity was achieved with the same accuracy, and the experimental setup was clear.
- 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.
(1) Insufficient explanation of the novelty differences between existing adaptive computing and multi-scale methods; (2) The key stability indicators during reinforcement learning training were not reported, and the potential impact of high variance in policy gradient estimation in medical image small sample high noise scenarios was not discussed. (3) Qualitative or quantitative analysis of failed cases was not provided. For medical applications, it is crucial to understand when a model may incorrectly focus on non lesion areas or excessively contract, resulting in loss of contextual information.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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?
Method novelty, experimental adequacy, clarity of charts, and other aspects
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes the Chained Perceptual Refinement (CPR) framework for high-resolution medical image classification. Starting from a low-resolution global view, this method progressively predicts the locations of interest and the scale of the cropped regions, extracting local evidence from the original high-resolution image and fusing it with a continuously updated perceptual state. The authors aim to preserve as much fine-grained information as possible, crucial for diagnosis, as possible while maintaining a relatively fixed backbone network input size and peak memory. Experiments on five public datasets and various backbone networks demonstrate that this method achieves a good trade-off between accuracy and computational efficiency.
- 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.
1.The paper addresses a highly practical problem: how to preserve fine-grained diagnostic information in medical images without performing dense, high-resolution inference.
2.The proposed stepwise refinement approach from global to local is intuitive and reasonable. Furthermore, the predicted location and scale are more suitable for medical images with significant resolution differences than fixed-size cropping.
3.Experiments cover five public datasets, multiple medical imaging modalities, and various backbone networks, providing more convincing validation than single-dataset validation.
4.The paper reports not only accuracy but also memory usage and GFLOPs, consistent with the method’s core efficiency claims.
5.Qualitative visualization and scale-controlled ablation experiments provide strong support for the method’s behavior.
- 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.
1.Regarding methodological novelty, the authors need to more clearly define the differences between this paper and existing adaptive vision/coarse-to-fine methods. The current work shares a clear conceptual connection with existing methods such as Adaptive NN, OverLock, and PAMIL in the core idea of “global overview followed by local region focus” or “dynamically driven region selection.” Therefore, based solely on the current manuscript, I cannot fully determine the key innovations of CPR and its irreplaceability compared to these works.
2.Ablation experiments are still incomplete. There is a lack of comparison with simpler alternatives, such as fixed coarse-to-fine scheduling, non-reinforcement learning pruning strategies, and deterministic attention mechanisms. Therefore, the necessity of PPO and the chained design cannot be fully confirmed at present.
3.There are some inconsistencies in the wording. For example, the abstract emphasizes predicting the crop center, while the method section defines (x_t, y_t) as the top-left coordinate; the descriptions of the state variables z_t and H_t are also inconsistent.
Wang, Yulin, et al. “Emulating human-like adaptive vision for efficient and flexible machine visual perception.” Nature Machine Intelligence (2025): 1-19. Lou, Meng, and Yizhou Yu. “Overlock: An overview-first-look-closely-next convnet with context-mixing dynamic kernels.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2025. Zheng, Tingting, Kui Jiang, and Hongxun Yao. “Dynamic policy-driven adaptive multi-instance learning for whole slide image classification.” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2024.
- 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 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
This work is promising overall, but I suggest the authors further enhance the transparency of the experiments and the details of the methodologies. In particular, I recommend clarifying the data partitioning protocol (whether it is patient-level partitioning in the applicable context), reporting the variance or statistical test results of multiple replicate experiments, adding clinically relevant metrics other than accuracy for class-imbalanced tasks, and adding control experiments against simpler non-reinforcement learning refinement strategies. Furthermore, adding a discussion of failure cases and methodological limitations would make the paper more complete.
- 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?
I believe this is a borderline, slightly positive submission. The paper addresses an important problem, its methodology is well-designed, and it demonstrates an attractive accuracy/efficiency tradeoff across multiple datasets. However, the current version is insufficiently novel compared to existing adaptive vision methods, lacking adequate implementation details, variance of repeated experiments, and stronger ablation studies. Therefore, the strength of the evidence is weaker than some of the paper’s claims. I currently lean towards a positive recommendation, primarily because the approach has practical value and broad experimental coverage, but I hope the authors will clarify issues regarding reproducibility and comparative fairness in their rebuttal.
- 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 rebuttal addresses my main concerns. The authors clarified the relationship to prior adaptive/coarse-to-fine methods, provided additional stability evidence with multiple random seeds, discussed concrete failure cases and limitations, and added stronger ablation/comparison results that better support the contribution of the chained policy design. While some implementation details could still be clearer, the paper’s practical significance and overall empirical support are now sufficient for acceptance.
Review #3
- Please describe the contribution of the paper
This paper introduces an image classification framework evaluated on high-resolution medical datasets. The core of the approach is an iterative, multi-pass inference strategy. In the initial pass, the full-resolution image is downsampled to a canonical resolution to extract global contextual features. Subsequently, a learned policy network predicts a bounding box to localize a salient region from the original high-resolution input. This region is then cropped, rescaled, and forwarded through the network to capture fine-grained local details. This process is repeated iteratively to progressively aggregate discriminative information. By maintaining a fixed canonical resolution throughout the passes, the model facilitates the processing of high-resolution imagery with a constant memory footprint. Experimental results across five datasets demonstrate that the proposed method achieves superior accuracy with lower computational overhead compared to existing baselines.
- 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 proposed method is supported by a thorough evaluation across five diverse datasets. The results consistently demonstrate that the framework achieves superior classification performance while maintaining significantly lower computational and memory overhead compared to traditional high-resolution processing techniques.
- 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 core concept of utilizing iterative local crops to extract fine-grained information is well-established in the literature, yet the manuscript lacks a comprehensive discussion of, and comparison with, relevant prior art. While the authors compare their work to AdaptiveNN, they overlook several foundational and contemporary models with similar architectural parallels:
- RA-CNN [a]: This model similarly employs a recurrent, iterative approach to propose variable-sized crops with intermediate supervision.
- SSPS/M2Former [b]: Recent work that also utilizes variable-sized patches for fine-grained recognition.
- Feature Accumulation Strategies: The proposed formulation closely resembles a recurrent network with a linear gate. Similar strategies using LSTMs for feature refinement have been proposed by Shroff et al. [c], and alternative aggregation methods, such as positional aggregation in TNet [d], already exist. Therefore, the authors should contextualize their contribution relative to these and other similar methods and ideally provide quantitative comparisons to justify the necessity of the new formulation.
- Evaluating against only one specialized model (AdaptiveNN) is insufficient for a paper claiming a superior “low-cost” approach. The authors should include standard iterative or hard-attention baselines to substantiate their claims of achieving a better accuracy-to-cost trade-off.
- To validate the efficacy of the policy network, the ablation study should include a configuration where crop boxes are randomly sampled rather than predicted. This is a critical baseline to demonstrate that the learned policy network is identifying salient regions and that the performance gains are not merely a byproduct of processing higher-resolution patches, regardless of their location.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The motivation behind this work, addressing the challenge of high-resolution image processing under strict resource constraints, is highly relevant to the medical imaging community. Furthermore, the empirical results demonstrate consistent performance gains across several benchmarks. However, the manuscript’s primary weakness is the lack of a thorough contextualization within the existing literature on coarse-to-fine iterative mechanisms. This paradigm has been extensively explored, and without a comprehensive comparison against both classical and modern “hard-attention” baselines, the technical novelty and comparative effectiveness of the proposed architecture remain unproven. Consequently, I am leaning toward a rejection at this stage. I would be open to revising my assessment only if the authors can provide a substantive rebuttal that demonstrates clear, unique contributions relative to some closely related methods.
- 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 proposed method is relevant, and it has been tested on multiple datasets to demonstrate its generalization capabilities. My main initial concern was the lack of discussion of previous similar methods, which has been addressed in the rebuttal. If the authors can include those in the final version of the paper, then I would like to increase my recommendation.
Author Feedback
We use W to denote Weakness and R to denote Reviewer.
(Joint Reply: R1-W1, R2-W1) Novelty and relation to adaptive/coarse-to-fine methods. We thank the reviewers for raising this point. CPR differs from prior adaptive, hard-attention, and coarse-to-fine methods in three key aspects: (1) it predicts both crop location and scale, rather than using fixed-size crops, predefined multi-scale inputs, or only location proposals; (2) it performs chained perceptual refinement, where each step updates a residual perceptual state instead of independently aggregating local crops; and (3) it targets high-resolution medical classification with fixed backbone input size and bounded memory. Compared with AdaptiveNN, CPR adds adaptive scale refinement; compared with OverLock, CPR revisits the original high-resolution image via adaptive crop-and-refine actions rather than context-mixing kernels; compared with PAMIL, CPR addresses image-level classification rather than WSI MIL.
(R1-W2) Stability, variance, and statistical reliability. We repeated CPR with three random seeds on COVID and BUSI datasets. CPR obtains 98.64 ± 0.17% on COVID and 95.16 ± 0.30% on BUSI, consistently outperforming the compared baselines. The small standard deviations indicate that CPR is robust across different runs, suggesting that the policy-gradient-based learning component does not introduce noticeable performance instability.
(R1-W3) Failure cases, clinical reliability, and limitations. We thank the reviewer for this important suggestion. Our qualitative inspection identifies two main failure modes. First, CPR may over-contract the crop when lesions are subtle, diffuse, or spread over a large region, which can lead to the loss of useful contextual information. Second, when visually salient non-lesion structures are present, the policy may attend to irrelevant regions and accumulate misleading evidence. These observations suggest that CPR is most effective when discriminative evidence is spatially localized, while diffuse, ambiguous, or weakly contrasted lesion patterns remain challenging. We will add this failure-case analysis to the revised manuscript.
(R2-W1) Novelty. Please see R1-W1. (R2-W2) Ablation completeness. To address the reviewer’s concern, we include additional ablation baselines on the BUSI dataset using the ViT-S model: Random Crop, Fixed Coarse-to-Fine, Deterministic Attention, and No-Chain CPR. These variants achieve 93.29%, 93.61%, 94.25%, and 94.25%, respectively, while the full CPR achieves the best performance of 94.89%. These results further validate the effectiveness of CPR.
(R2-W3) inconsistencies in wording. We will thoroughly check the manuscript revise the inconsistent wording to ensure clarity and consistency in the revised version.
(R3-W1) Similar Architectural Baselines. We thank the reviewer for this helpful suggestion. To address this concern, we additionally compare CPR with four related architectural baselines: (1) RA-CNN, (2) M2Former, (3) Focus Longer to See Better, and (4) TNet. On BUSI, these methods achieve 88.82%, 93.61%, 90.10%, and 94.25%, respectively, while CPR achieves the best performance of 94.89%. On COVID, they achieve 98.25%, 95.71%, 95.59%, and 98.37%, respectively, while CPR further improves the performance to 98.67%. These results demonstrate that CPR consistently outperforms these baselines, further validating the effectiveness of our proposed method.
(R3-W2) Policy network ablation. We thank the reviewer for this insightful suggestion. To address this concern, we report two additional baselines on the BUSI dataset using the ViT-S model: CPR with Random Crop and CPR with Deterministic Attention. These baselines achieve 93.29% and 94.25%, respectively, while the full CPR achieves 94.89%. These results indicate that the learned policy network provides more effective region selection than random cropping or deterministic attention, thereby contributing to the overall performance improvement.
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.
This manuscript has received mixed reviews (scores: 4, 4, 3), with major concerns centered on insufficient novelty differentiation from existing adaptive vision methods, incomplete ablation studies, and lack of comparison with relevant coarse-to-fine baselines. The authors are suggest to address these limitations in the rebuttal phase.
- 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.
After rebuttal, all reviewers decide to accept the paper, the author successfully adderss the problems , and I recommend the acceptance of the paper.
Meta-review #2
- 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 paper addresses an important and practical problem in high-resolution medical image classification, aiming to improve the balance between diagnostic performance and computational efficiency. The initial reviews raised concerns about novelty, comparison with related adaptive/coarse-to-fine methods, and the completeness of ablation studies. The authors’ rebuttal provides reasonable clarification addressing these main concerns. Importantly, the reviewer opinions converged toward acceptance after the rebuttal, indicating agreement that the major concerns were sufficiently addressed. While the method remains somewhat incremental and some implementation details could still be improved, the overall empirical results across multiple datasets are positive. I therefore recommend acceptance.
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.
After considering the reviews and rebuttal discussion, I recommend acceptance. The paper addresses an important and practical challenge in medical imaging: preserving fine-grained diagnostic information in high-resolution images under limited computational budgets.
The proposed coarse-to-fine refinement framework is well motivated and demonstrates a strong balance between accuracy, memory efficiency, and computational cost across multiple datasets, modalities, and backbone architectures. In particular, the chained refinement mechanism and adaptive scale-aware region selection provide a clinically meaningful alternative to fixed-resolution or fixed-crop strategies.
While some concerns regarding novelty relative to prior adaptive vision and coarse-to-fine methods were initially raised, the rebuttal satisfactorily clarified the distinctions from related work and strengthened the empirical support through additional analyses, ablations, and discussions of stability and failure cases. Overall, the paper presents a technically sound and practically valuable contribution with sufficiently broad experimental validation for acceptance at MICCAI.