List of Papers Browse by Subject Areas Author List
Abstract
Whole Slide Image (WSI) analysis in digital pathology requires continual learning (CL) to adapt to growing data streams without catastrophic forgetting. Current CL methods for WSIs face challenges from large sample sizes and strict patient privacy constraints, risking either data leakage or limited performance on visually similar categories.
Existing methods suffer from inter-task class confusion due to insufficient discriminative supervision, and an inability to update prior class representations.
To address these, we propose AlignMIL-NS with two innovations: (1) Prototype-Class Feature Alignment (PCFA), which enforces within-task alignment between class and visual features while promoting inter-task discrimination; and (2) Null Space Class Feature Update (NS-CFU), which refines previous task representations within their task-specific null spaces to mitigate interference from new tasks.
On a challenging five-dataset WSI benchmark, AlignMIL-NS achieves 89.1\% AACC on both task orders, outperforming all rehearsal-free methods and matching rehearsal-based approaches without storing any patient data.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2958_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{LiXia_Aligning_MICCAI2026,
author = { Li, Xianrui AND Xiao, Kaiwen AND Chan, Antoni B.},
title = { { Aligning Prototypes and Updating Null Spaces for Continual WSI Learning } },
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 introduces a rehearsal-free continual learning framework designed for digital pathology. It addresses the catastrophic forgetting that occurs when MIL models are sequentially trained on new WSI 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.
1.Unlike conventional methods that typically freeze old representations to avoid forgetting, the proposed method allows the model to continuously refine previous-task class representations. This is achieved by constraining updates to the task-specific null space. 2.The proposed method addresses the challenge of inter-task class confusion by enforcing alignment between class features and visual prototypes within a single task. Because historical visual features are inaccessible in a privacy-preserving setting, PCFA uses prototype features as a proxy to promote cross-task separability, which is particularly effective for distinguishing between visually similar disease categories across sequential data streams.
- 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 stability of this framework is highly dependent on the fine-tuning of hyperparameters. To balance model plasticity (the ability to acquire new knowledge) with stability (the ability to preserve previously learned knowledge), users must configure specific settings for the null space threshold gamma, which serves as a margin hyperparameter, as well as the weights for various loss functions. The generalizability of these hyperparameters remains to be further validated when encountering different pathological tasks or more complex, multi-center data streams.
- 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 recommendation is primarily based on the framework’s strong clinical utility and its ability to achieve performance comparable with rehearsal-based methods without storing raw patient data, thus strictly respecting medical privacy constraints. With competitive performance across a rigorous five-dataset benchmark, the proposed method proves highly effective at mitigating catastrophic forgetting by addressing the visual similarity and inter-task interference common in WSI analysis by enforcing discriminative supervision and allowing non-destructive updates to old knowledge. Additionally, the architecture is remarkably efficient. However, there is a notable dependency on the fine-tuning of hyperparameters, which may hinder its robustness when applied to varying pathological tasks or multi-center data streams.
- Reviewer confidence
Very confident (4)
- [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
The authors propose AlignMIL-NS to solve the challenges in the current continual learning methods, large sample sizes and strict patient privacy. The innovations are: (1) Prototype-Class Feature Alignment (PCFA), which enforces within-task alignment between class and visual features while promoting inter-task discrimination; and (2) Null Space Class Feature Update (NS-CFU), which refines previous task representations within their task-specific null spaces.
- 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 identified the critical issues in the previous methods, the lack of explicit cross-task discriminability constraints and the inability to update previous class representations. Two proposed components (PCFA and NS-CFU) are conceptually aligned with the identified problems. 2.The paper demonstrates the robustness to the task order, and includes the ablations of PCFA and NS-CFU, respectively.
- 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 Prototype-based MIL, what is P_i^x when constructing prototype features? 2.Fig1 makes me confused on which one is QPMIL-VL heatmaps and which one is AlignMIL-NS heatmaps. 3.If the class feature is either learned by a learnable vector or updated in the null space, then what is the point of a fixed language encoder in class feature? 4.The class feature is driven by a fixed language encoder and a text template, which is initialized in the text space. But in the null space class feature update, the class feature is updated in the null space of visual features. This only works under the assumption that the text space and visual space are aligned, but there is no guarantee on the alignment. 5.In the ablation study, the authors should also include BWT and IM to measure forgetting and plasticity because the authors state that “PCFA loss can effectively addressing catastrophic forgetting”
- 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?
Same as weakness
- Reviewer confidence
Somewhat confident (2)
- [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 #3
- Please describe the contribution of the paper
This paper focuses on the problem of pathology slide classification using a continual learning setting, specifically prototype-based multiple instance learning. In this setting, tasks are learned sequentially, and the feature extractor evolves over time. The feature embeddings are updated and class prototypes from earlier tasks are typically frozen. This leads to two problems: (i) a mismatch between outdated prototypes and current embeddings, and (ii) misalignment between learned prototypes and the true class feature distributions.
To address these issues, the paper proposes two main components. First, an alignment step that forces prototypes to closely match their corresponding class feature distributions (within-task alignment) while also explicitly separating features across different tasks (cross-task discriminability). This results in tighter class clusters and improved separation between classes from different tasks. Second, they introduce null space update that moves prototypes from previous tasks instead of freezing them. The updates are constrained to lie in the null space of the feature subspace associated with each previous task, i.e., the new information can be incorporated without altering the original task-specific representations. This allows prototypes to adapt to the evolving feature space while avoiding interference with previously learned knowledge.
The paper is well supported experimentally using five WSI datasets, ablation studies, and comparisons against 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 paper identifies known issues with MIL and explains them well. Then, it proposed two natural solutions to handle these issues. It provides geometric interpretation of their solutions. The experimental results are also solid. The improvements are as expected.
- 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.
I am not sure if there is any major weakness. Perhaps the paper assumes that a reader is well versed in continual learning but that is expected in such conferences.
- 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.
(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?
I am not an expert in this field but based on my understanding this paper is definitely suggesting some natural and significant improvements to handle known problems in MIL.
- 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
Author Feedback
We thank all reviewers for their constructive feedback. We address the major concerns below.
1.Hyperparameter Sensitivity (Meta-review, R1)
We tested $\gamma$ across different values. Although model performance is affected by this hyperparameter, it consistently brings positive gains over the baseline (AACC 77.3% without NS-CFU) within the range of [0, 0.6]. Moreover, the null space constraint (Sec. 3.3) guarantees $\langle f_q^j, f_c^j + \phi_j \alpha_j^\top \rangle = \langle f_q^j, f_c^j \rangle$, meaning $\gamma$ only affects cross-task discrimination and cannot degrade within-task performance. Loss weights ($\lambda_M$=0.5, $\lambda_{\text{PCFA}}$=1.0) remained fixed across both task orders and all five folds without per-task tuning.
2.Role of the Fixed Language Encoder (R2-Q3, Meta-review)
Fine-tuning the lightweight $\mathbf{v}_c$ requires far fewer parameters than fine-tuning $E_l$ (0.031M vs. 0.479M, Tab. 5). More importantly, in the CL setting, fine-tuning $E_l$ would alter all class features simultaneously, whereas updating $\mathbf{v}_c$ only modifies the target class feature, avoiding interference with previously learned classes.
3.Text-Visual Alignment (R2-Q4, Meta-review)
Our framework operates under CONCH (Sec. 4.1), where both patch features ($E_v$) and text templates ($E_l$) are encoded by the same vision-language model, which is explicitly trained with contrastive learning to align text and visual embeddings. Therefore, the alignment between text and visual spaces is an architectural property, not an unsupported assumption. We will clarify this more explicitly in the manuscript.
4.Ablation Metrics (R2-Q5)
AACC is the primary metric in class-incremental learning. Since the joint-training upper bound (90.3%) is the same across all configurations, a higher AACC directly indicates better forgetting resistance. The full model achieves BWT=-4.0 vs. -15.1 for QPMIL-VL (Tab. 2), confirming that both PCFA and NS-CFU contribute to reduced forgetting.
5.Notation and Figure Clarity (R2-Q1, R2-Q2)
We acknowledge that $P_i^x$ in Eq. 1 is not clearly defined. It refers to the keys $k_i$ in the prototype pool as described in the QPMIL-VL formulation. We will fix this notation. Regarding Fig. 1, we apologize for the confusion: the current figure only shows the QPMIL-VL heatmap. We will add the AlignMIL-NS heatmap and improve the in-figure annotations.
6.Accessibility (R3)
We will improve the exposition to be more accessible for readers less familiar with continual learning.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 consensus is positive, with two Weak Accepts and one Accept. The work addresses an important and clinically relevant problem: continual WSI learning under privacy constraints. The methodological contribution is reasonably clear and targeted, and the results appear competitive across a challenging multi-dataset benchmark. The main concerns are not fatal but should be clarified in rebuttal or revision. In particular, the technical validity of updating text-derived class representations in a visual null space needs stronger explanation, and the sensitivity of key hyperparameters should be better supported.
