List of Papers Browse by Subject Areas Author List
Abstract
Multiple Instance Learning (MIL) has emerged as the dominant paradigm for computational Whole Slide Images (WSIs) analysis. To capture richer diag-nostic cues while maintaining computational efficiency, recent MIL methods have adopted coarse-to-fine sampling strategies that use low-magnification overviews to guide high-resolution patch selection. However, the unidirection-al nature of such strategies prevents the model from correcting suboptimal re-gion selection, leading to irreversible loss of critical diagnostic information. To address these challenges, we propose GFR-MIL, a novel Glance-Focus-Reflect framework where the Reflect module provides top-down feedback from high-resolution fine-grained features to recalibrate low-resolution global representations, establishing bidirectional cross-scale interaction for MIL-based WSI analysis. This iterative closed-loop mechanism emulates the cogni-tive workflow of experienced pathologists, who rapidly glance at the global tissue context, focus on suspicious regions, and reflectively revise initial judgments. To further address information loss from unsampled regions, we in-troduce a Recycling Mechanism and a memory masking strategy, ensuring comprehensive slide-level supervision throughout training. Extensive experi-ments across 7 public cohorts demonstrate that GFR-MIL consistently outper-forms state-of-the-art methods on both cancer subtyping and prognosis predic-tion tasks, with thorough ablation studies validating the contribution of each proposed component.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2838_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/longger-z/GFR-MIL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{QiMin_GFRMIL_MICCAI2026,
author = { Qi, Mingxin AND Li, Hongyu AND Lu, Yinuo AND Mu, Wei},
title = { { GFR-MIL: Glance–Focus–Reflect Based Multiple Instance Learning for Multi-Scale Whole Slide Image Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS ActiveX VT_ERROR: },
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes GFR-MIL, a model that mimics the image-reading patterns of experts and is designed according to the Glance-Focus-Reflect cognitive paradigm. It achieves patch aggregation and WSI analysis through bidirectional closed-loop interaction between low-resolution global context and high-resolution local fine-grained features. The paper evaluates the model’s performance on six public datasets and presents interpretability heatmaps.
- 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 authors have modeled the image-reviewing process of pathologists, with a meticulously designed framework that provides a counterintuitive discussion of the necessity of each component. 2) The authors conducted a comprehensive evaluation of their method against nine other approaches across six datasets, covering cancer prognosis prediction and cancer subtyping tasks; overall, their method is competitive.
- 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) Although the authors conducted extensive comparisons of model performance, the improvement in subtype prediction performance on TCGA-NSCLC relative to the baseline appears to be minimal, and the authors did not report on statistical significance. 2) Unlike other datasets, confidence intervals were not reported when comparing with BRCAS. 3) During the sampling phase of the Glance Module, the authors claim to select the top-k% instances, but they do not report the value of k nor analyze the impact of varying k on model sensitivity. 4)While the paper repeatedly emphasizes computational efficiency through coarse-to-fine selection and claims that GFR-MIL achieves strong performance while attending to only 30% of slide regions, the actual end-to-end efficiency remains insufficiently substantiated. In particular, the framework still requires dual-resolution preprocessing and feature extraction, iterative stacked GFRBlocks, cross-scale feedback through the Reflect module, recycling of unselected regions, and multi-round feature aggregation, all of which may introduce non-trivial computational and memory overhead.
- 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?
The proposed method exhibits a certain degree of novelty, but the key parameters of the model have not been explained. Performance improvements in the classification task are modest.The authors claim to have balanced efficiency and accuracy; however, the paper’s persuasiveness would be enhanced if they reported convincing efficiency metrics such as GPU memory usage and FLOPs.
- Reviewer confidence
Very confident (4)
- [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 have provided a targeted and largely satisfactory response to my four concerns:
NSCLC improvement (Q1): They acknowledge that TCGA‑NSCLC subtyping is a near‑saturated task relying on few discriminative patches, and redirect attention to the prognosis tasks (five datasets) where GFR‑MIL achieves consistent and notable C‑index gains. They also report a 44% reduction in preprocessing cost (27.5s vs 48.9s for ABMIL‑Gated). This is an acceptable clarification — the method’s value is clearer on harder tasks, while remaining efficient on saturated ones.
BRACS confidence intervals (Q2): They correctly note that BRACS provides official fixed train/val/test splits, and they followed this standard protocol. Under such fixed splits, reporting standard deviations is not standard practice. This resolves my concern.
Top‑k% sampling ratio (Q3): The authors explicitly state that the two GFRBlocks use k = 20% and 10% (total ~30% of patches), referencing [28]. They also argue that the closed‑loop Reflect module recovers missed regions in round 2, making the framework architecturally robust to k, supported by SOTA performance across 7 datasets. This is a reasonable justification, though an ablation on k would have been stronger. I accept the explanation.
Computational efficiency (Q4): This is addressed convincingly. The authors now provide concrete numbers (to be added to the final paper)
The rebuttal has successfully resolved my major concerns, particularly the critical issue of computational efficiency. The method remains novel and well motivated, and the empirical evaluation is solid. I maintain my original score of Weak Accept.
Review #2
- Please describe the contribution of the paper
The main contribution of this work lies in the formulation of an iterative cross-scale MIL framework for WSIs, where global and local information are not simply fused once, but are progressively coordinated through repeated rounds of selection, local modeling, and feedback refinement. Compared with conventional MIL pipelines that rely on one-shot instance selection or flat aggregation, the proposed method introduces a more structured diagnostic reasoning process and provides several task-specific designs, including gated regional sampling, recycling of unselected information, feedback-based low-resolution updating, and history-aware region suppression.
- 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 proposed GFR-MIL framework is well motivated by the diagnostic workflow of pathologists and provides a clear and intuitive coarse-to-fine reasoning paradigm. 2.The proposed gated sampling, entropy regularization, and recycling mechanism together form a relatively complete sparse selection pipeline that is both practical and technically sound. 3.The iterative Reflect module with memory constraint introduces a meaningful multi-round refinement strategy, which is conceptually appealing for WSI analysis and has the potential to improve representation completeness.
- 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.In Fig. 1, the number of stacked GFRBlocks is denoted as N, but the manuscript does not clearly specify its value. It is also unclear whether the parameter D introduced later refers to the number of GFRBlocks. This point should be clarified, as it appears to be a key architectural hyperparameter. 2.The paper does not provide any analysis of the computational cost, such as FLOPs, inference time, or GPU memory consumption. Since the proposed GFRBlocks appear to introduce nontrivial overhead, and the dual-resolution design may further increase both computational complexity and inference latency in practical deployment, it would be important to analyze the trade-off between performance and efficiency. 3.The visualization in Fig. 2 is currently not sufficiently clear. In particular, the semantic meaning of different colors (e.g., red vs. blue) is not well explained, making it difficult to understand what exactly the figure is intended to demonstrate. More interpretable visualizations and clearer figure captions would strengthen the paper. 4.Since the method is based on multi-round Glance–Focus–Reflect reasoning, it would be helpful to analyze whether different blocks attend to different types of regions. For example, do earlier blocks focus on the most salient lesions, while later blocks identify complementary or previously overlooked regions? Such analysis is currently missing.
- 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.
(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?
This paper proposes a novel GFR-MIL framework that models iterative cross-scale interaction for WSI analysis. The Glance–Focus–Reflect paradigm is conceptually appealing and the experimental results show consistent improvements over strong baselines. However, several aspects are insufficiently validated. In particular, computational cost, and efficiency–performance trade-off are not analyzed. In addition, the claims regarding the Reflect module and iterative reasoning are not fully supported by detailed analysis or clear visualizations. Overall, despite these limitations, the idea is novel and promising. I lean towards a weak accept, while encouraging stronger computational cost analysis.
- Reviewer confidence
Very confident (4)
- [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 #3
- Please describe the contribution of the paper
This paper proposes GFR-MIL, a multi-scale MIL framework for whole slide image analysis. The main idea is to move beyond the standard one-way coarse-to-fine pipeline by introducing a Glance–Focus–Reflect design, where high-resolution features can in turn refine the low-resolution representation through feedback. The paper also introduces a recycling mechanism to reduce the information loss caused by hard region selection.
- 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 paper studies an important problem in WSI analysis, namely the limitation of irreversible region selection in existing coarse-to-fine MIL methods. The overall motivation is reasonable, and the proposed framework is intuitive and easy to follow. Another positive aspect is that the method is evaluated on both subtyping and survival prediction tasks across several public datasets, and the reported performance is generally strong.
- 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 main concerns are mostly on the experimental side. First, the baseline comparison is not fully convincing, since the paper does not clearly compare with the most relevant coarse-to-fine or hierarchical multi-scale methods. Second, the BRACS results are reported without standard deviation, which makes it hard to judge the stability of the method. Third, although the paper emphasizes efficiency, this aspect is not quantitatively evaluated with metrics such as runtime, memory, FLOPs, or processed patch ratio. Fourth, for the survival task, the evaluation is limited to C-index only, and additional survival-related analysis would make the results more convincing. Finally, the visualization section only presents the proposed method, without qualitative comparisons to other baselines, so the interpretability advantage is not fully demonstrated.
- 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.
(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 main concern of this paper is that the experiments are insufficient to prove the effectiveness of its method.
- 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.
Since my concerns are not addressed by the rebuttal, my final opinion is reject.
Author Feedback
We thank all reviewers for the constructive feedback, and code will be released upon acceptance. Q1 for R2&Meta: NSCLC Improvement. Subtyping on TCGA-NSCLC is a near-saturated task as it depends on a few discriminative patches[26]. GFR-MIL matches top performance while reducing preprocessing cost by 44% (27.5s vs 48.9s for ABMIL-Gated, median 15,135 patches @20x). On prognosis tasks requiring global WSI context [27], GFR-MIL consistently outperforms all competitors across five datasets with notable C-index gains (Table 1). GFR-MIL thus offers efficiency on saturated tasks and accuracy on challenging ones. [26] IEEE TMI, vol. 44, no. 9, pp. 3756-3769.[27] CVPR, 2025, pp. 5144-5153.Q2 for R2, 8&Meta: BRACS Confidence Intervals. BRACS provides official fixed train/val/test splits (Sec 3.1), we follow this standard protocol for direct comparison with prior work. Q3 for R2: Top-k% Sampling Ratio. The two GFRBlocks use k = 20% and 10% (Sec 3.2), selecting ~30% of patches in total, consistent with [28]. Notably, the closed-loop Reflect module recovers regions missed in round 1 during round 2, making GFR-MIL architecturally robust to k. This is supported by SOTA performance across 7 datasets with widely varying patch counts (Tables 1, 2). [28] arXiv:2603.00667 Q4 for R2, 3, 8&Meta : Computational Efficiency and Trade-Off Recycling Mechanism is training-only to ensure gradient flow (Sec 2.2). At inference, GFR-MIL processes only the 2.5× WSI and ~30% of 20× patches. On TCGA-KIRC (median 19,187 patches @ 20×), preprocessing (dual-resolution feature extraction) drops from 51.3s (full 20×) to 30.4s per slide, and the full pipeline (stacked GFRBlocks, Reflect feedback, and multi-round aggregation) costs 85.7G FLOPs, 0.55GB memory, and 11.3ms inference, versus 189.5G/0.93GB/14.5ms (TransMIL) and 571.7G/2.0GB/29.8ms (RRT). Similar trends hold across all 7 datasets. These omitted details will be added. Q5 for R3: Parameters N and D. N in Fig. 1 denotes the number of stacked GFRBlocks, equivalent to D in Sec 3.2, with a value of 2.We will unify this notation in the final version. Q6 for R3: Fig. 2 Visualization Clarity. The red highlighted areas denote the high-attention regions identified by GFR-MIL, while the blue areas indicate the low-attention regions. We will clarify this in the caption. Q7 for R3: Multi-Round Attention Behavior. The two GFRBlocks are functionally complementary by design. The first selects salient top-ranked regions, and the memory mask (Sec 2.4) then suppresses them, forcing the second to find overlooked yet discriminative ones (0% spatial overlap). Ablation confirms each contributes (Table 3). We will color-code Fig.2 by round, with the first round in black boxes and the second in green. Per-round histological profiling is left for future interpretability work. Q8 for R8&Meta: Coarse-to-Fine Baselines. We addresse coarse-to-fine baselines from two angles. (1) RLogist [13], a representative coarse-to-fine method, reports in its own paper that it underperforms high-resolution MIL baselines. We therefore prioritized comparisons with the latest and strongest high-resolution methods [19, 20], where GFR-MIL achieves SOTA across all 7 cohorts. (2) In Table 3, replacing our bidirectional Reflect with a unidirectional Gumbel-Softmax reduces GFR-MIL to a standard coarse-to-fine pipeline, and the resulting drop confirms our design’s advantage. Q9 for R8: Survival Evaluation Metrics. C-index is a standard metric in computational pathology survival analysis, widely adopted in recent works [20, 27]. GFR-MIL ranks first across all 5 distinct cancer cohorts, validating its superiority. Further KM analyses will be conducted in future. Q10 for R8: Qualitative Baseline Comparison. GFR-MIL achieves unique interpretability. Unlike conventional MIL with diffuse whole-slide attention heatmaps [2], it outputs discrete localized boxes for per-round selected regions (Fig. 2). Qualitative baseline heatmap comparison is left for future work.
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 paper has a coherent and potentially useful idea for WSI MIL, and two reviewers lean toward acceptance. However, the acceptance case depends on whether the authors can substantiate the experimental claims. The method is presented as efficient and more complete than previous coarse-to-fine MIL approaches, but the current evidence does not sufficiently support this. In particular, the lack of runtime, memory, FLOPs, patch ratio, and end-to-end preprocessing cost weakens the central efficiency claim. The experimental comparison is also not fully convincing unless the authors include or clearly justify the absence of relevant hierarchical/coarse-to-fine baselines. The reported improvements appear modest in some settings, and missing confidence intervals or statistical significance further limits confidence.
- 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 proposed Glance–Focus–Reflect paradigm is conceptually appealing to me, and the empirical results demonstrate consistent improvements over robust baselines.
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.
This paper received two accept and one reject after rebuttal. I recommend accepting this paper. The overall motivation and idea is recognized by the reviewers and the major concerns are about experimental evaluation. The rebuttal addressed some of the concerns, while someother are difficult to address without additional experiments. The proposed method is evaluated on seven datasets, so I think the experiments are fairly comprehensive.
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 paper received mixed reviews in the first round, and the authors were invited to submit a rebuttal. The authors submitted a rebuttal to address the raised points. The two positive reviewers remained positive, noting that the authors’ answers addressed the raised points. On the other hand, the negative reviewer highlighted that their comment had not been properly addressed. The AC read the paper, reviews, rebuttal, and final rating and agrees that the paper has merit and it will be interesting for the community. The authors are encouraged to include their responses to the camera-ready version of the paper.
